hermes-agent/TODO.md

Hermes Agent - Future Improvements

---

What We Already Have (for reference)

42+ tools across 12 toolsets: web (search, extract), terminal + process management, file ops (read, write, patch, search), vision, MoA reasoning, image gen, browser (10 tools via Browserbase), skills (41 skills), cronjobs, RL training (10 tools via Tinker-Atropos), TTS, cross-channel messaging.

4 platform adapters: Telegram, Discord, WhatsApp, Slack -- all with typing indicators, image/voice auto-analysis, dangerous command approval, interrupt support, background process watchers.

Other: Context compression, context files (SOUL.md, AGENTS.md), session JSONL transcripts, batch runner with toolset distributions, 13 personalities, DM pairing auth, PTY mode, model metadata caching.

---

1. Subagent Architecture (Context Isolation) 🎯

Status: Not started Priority: High -- this is foundational for scaling to complex tasks

The main agent becomes an orchestrator that delegates context-heavy tasks to subagents with isolated context. Each subagent returns a summary, keeping the orchestrator's context clean.

Core design:

• delegate_task(goal, context, toolsets=[]) tool -- spawns a fresh AIAgent with its own conversation, limited toolset, and task-specific system prompt
• Parent passes a goal string + optional context blob; child returns a structured summary
• Configurable depth limit (e.g., max 2 levels) to prevent runaway recursion
• Subagent inherits the same terminal/browser session (task_id) but gets a fresh message history

What other agents do:

• OpenClaw: sessions_spawn + subagents tool with list/kill/steer actions, depth limits, rate limiting. Cross-session agent-to-agent coordination via sessions_send.
• Codex: spawn_agent / send_input / close_agent / wait_for_agent with configurable timeouts. Thread manager for concurrent agents.
• Cline: Up to 5 parallel subagents per invocation. Subagents get restricted tool access (read, list, search, bash, skill, attempt). Progress tracking with stats (tool calls, tokens, cost, context usage).
• OpenCode: TaskTool creates subagent sessions with permission inheritance. Resumable tasks via task_id. Parent-child session relationships.

Our approach:

• Start with a single delegate_task tool (like Cline's model -- simple, bounded)
• Subagent gets: goal, context excerpt, restricted toolset, fresh conversation
• Returns: summary string (success/failure + key findings + any file paths created)
• Track active subagents so parent can reference them; limit concurrency to 3
• Later: add send_input for interactive subagent steering (Codex-style)
• Later: cross-session coordination for gateway (OpenClaw-style sessions_send)

---

2. Planning & Task Management 📋

Status: Not started Priority: High -- every serious agent has this now

A todo tool the agent uses to decompose complex tasks, track progress, and recover from failures. Must be cache-friendly -- no system prompt mutation, no injected messages that invalidate the KV cache prefix.

What other agents do:

• Cursor (Claude): TodoWrite tool. Flat list, 4 states, merge flag. Todos injected into context every turn. Effective but cache-hostile -- mutating the system prompt on every turn invalidates the entire prefix cache for all prior tokens.
• OpenCode: todowrite / todoread as separate tools. State lives server-side. Agent reads it back when needed. Cache-friendly -- no context mutation; the todo state appears only as tool call/response pairs in the normal conversation flow.
• Cline: focus_chain tool for task management with sequential focus
• OpenClaw: /mesh <goal> command auto-plans and runs multi-step workflows

The caching problem

With OpenRouter / vLLM / any prefix-caching provider, the prompt is cached from the start of the sequence. If we inject a changing todo list into the system prompt (or anywhere before the most recent messages), we invalidate the cache for the entire conversation. On a 100K-token conversation, this means re-processing the full prefix on every turn instead of just the new tokens. That's a massive cost and latency penalty.

Rule: Never modify anything before the latest assistant/user turn. Anything we add must go at the end of the conversation (as a tool response or appended to the latest message), or live entirely inside tool call/response pairs that the agent initiated.

Design: Pure tool-based, no context mutation

Two tools: todo_write and todo_read

todo_write -- create or update the task list:

Parameters:
  todos: [{id, content, status}]   # the items to write
  merge: bool (default false)      # false = replace list, true = update by id

Returns: the full current todo list (so the agent immediately sees the result)

todo_read -- retrieve the current task list:

Parameters: (none)

Returns: the full current todo list, plus metadata (items completed, items remaining, tool calls since last update)

This is how OpenCode does it and it's the right call. The agent's own tool call history is the "memory" -- the todo_write response containing the full list is right there in the conversation. No injection needed.

Item schema:

{
  "id": "1",           # unique string identifier
  "content": "...",    # description of the task
  "status": "pending"  # one of: pending, in_progress, completed, cancelled
}

No priority field. Order in the list is the priority.

Only 4 states. No "failed" or "blocked" -- if something fails, cancel it and add a revised item.

Where the behavior rules live: the tool description

Instead of adding instructions to the system prompt, we put all the usage guidance directly in the tool description. This is part of the tool schema, which is sent once at the start and is cached perfectly -- it never changes mid-conversation.

The todo_write tool description teaches the agent everything:

Manage your task list for the current session. Use this to plan and track
multi-step work.

When to use: Complex tasks with 3+ steps, when the user gives multiple tasks,
or when you need to plan before acting. Skip for simple single-step requests.

Behavior:
- Set merge=false to create a fresh plan (replaces existing todos)
- Set merge=true to update status or add follow-up items
- Mark the first item in_progress immediately and start working on it
- Only keep ONE item in_progress at a time
- Mark items completed as soon as you finish them
- If something fails, mark it cancelled and add a revised item
- Don't add "test" or "verify" items unless the user asks for them
- Update todos silently alongside your other work

Returns the full current list after every write.

This is ~150 tokens in the tool schema. It's static, cached, and teaches the agent the same behavioral rules without touching the system prompt.

How it survives context compression

The state lives server-side on the AIAgent instance (in-memory dict). The conversation history just has the tool call/response pairs as a trail of what happened.

On context compression, re-inject the current todo state. Compression already invalidates the cache (the middle of the conversation is being rewritten anyway), so there's zero additional cache cost to appending the todo state once at that moment. When run_agent.py runs compression, if _todo_state is non-empty, it appends a synthetic tool response at the end of the compressed history:

[System: Your current task list was preserved across context compression]
- [x] 1. Set up project structure (completed)
- [>] 2. Implement the search endpoint (in_progress)
- [ ] 3. Add error handling (pending)

This is the one place we do inject -- but only on compression events, which are rare (once every ~50-100 turns). The cache was already blown by compression itself, so it costs nothing extra.

The agent does NOT need to "know" to call todo_read after compression -- it just sees its plan right there in the history. todo_read still exists as a tool for any time the agent wants to double-check, but it's not load-bearing for the compression case.

Progress checkpoints (cache-friendly approach)

Instead of injecting system messages (which mutate the context), we piggyback on existing tool responses. When any tool returns its result, and the tool call counter has crossed a threshold, we append a one-line note to that tool's response:

[10+ tool calls since last todo update -- consider calling todo_read to review your plan]

This is a tiny addition to a response that's already at the end of the conversation -- zero cache impact on previous turns. It costs ~20 tokens and only appears once per threshold crossing.

Similarly, on tool errors, we can append to the error response:

[This error may affect your plan -- call todo_read to review]

These hints go in the tool response, not in a new system message. The agent processes them naturally as part of reading the tool output.

Server-side state

# On AIAgent class
_todo_state: Dict[str, List[dict]]  # session_key -> list of todo items
_todo_tool_calls_since_update: int  # counter for checkpoint nudges

• todo_write updates _todo_state and resets the counter
• todo_read reads _todo_state and resets the counter
• Every handle_function_call increments the counter
• When counter > threshold (default 10), the next tool response gets the checkpoint hint appended, then the hint flag is cleared until the next threshold crossing

Summary: what we took from Cursor, what we changed

Aspect	Cursor's approach	Our approach
State visibility	Injected into context every turn	Tool call/response pairs in normal conversation flow
Behavioral rules	System prompt instructions	Tool description (static, cached)
Checkpoints	Injected system messages	One-line hint appended to tool responses
Cache impact	High (prefix invalidated every turn)	Near-zero (only re-injects on compression, which already blows cache)
State persistence	Context injection survives compression	Server-side dict; re-injected once on compression events
Core UX	Identical	Identical (same flat list, same 4 states, same merge semantics)

Files: tools/todo_tool.py (tool implementation), integration in run_agent.py (state dict + checkpoint counter)

---

3. Dynamic Skills Expansion 📚

Status: Partially implemented (41 read-only skills exist) Priority: Medium

Extend the skills system so the agent can create, edit, and delete skills at runtime. Skill acquisition from successful task patterns.

What other agents do:

• OpenClaw: ClawHub registry -- bundled, managed, and workspace skills with install gating. Agent can auto-search and pull skills from a remote hub.
• OpenCode: SKILL.md format with URL download support via Discovery.pull. Compatible with Claude Code's skill format.
• Pi: Skills as npm-publishable packages. Prompt templates and themes alongside skills.
• Cline: Global + project-level skill directories. use_skill tool + new_rule tool for creating rules.

Our approach:

• Add skill_create, skill_edit, skill_delete actions to the existing skill_view tool (or a new skill_manage tool)
• New skills saved to ~/.hermes/skills/ (user-created, separate from bundled)
• SKILL.md format stays the same (YAML frontmatter + markdown body)
• Skill acquisition: after a successful multi-step task, offer to save the approach as a new skill (agent-initiated, user-confirmed)
• Later: skill chaining with dependency graphs, parameterized templates
• Later: remote skill registry (like ClawHub) for community-shared skills

---

4. Interactive Clarifying Questions ❓

Status: Not started Priority: Medium

Allow the agent to present structured choices to the user when it needs clarification. Rich terminal UI in CLI mode, graceful fallback on messaging platforms.

What other agents do:

• Codex: request_user_input tool for open-ended questions
• Cline: ask_followup_question tool with structured options
• OpenCode: question tool for asking the user

Our approach:

• clarify tool with parameters: question (string), choices (list of up to 6 strings), allow_freetext (bool)
• CLI mode: Rich-powered selection UI (arrow keys + number shortcuts)
• Gateway/messaging mode: numbered list with "reply with number or type your answer"
• Returns the user's selection as a string
• Agent can use this before starting expensive operations ("Which approach do you prefer?")

File: tools/clarify_tool.py -- presentation layer differs per platform, core logic is simple

---

5. Memory System 🧠

Status: Not started Priority: High -- biggest gap vs. OpenClaw and Dash

Persistent memory that survives across sessions. The agent remembers what it learned, who the user is, and what worked before.

What other agents do:

• OpenClaw: 78+ file memory subsystem. SQLite + sqlite-vec for vector search. Embeddings via OpenAI/Voyage/Gemini. Hybrid search (vector + keyword). MMR for diversity. Temporal decay. File watcher for auto-indexing. Session-scoped memory with citations.
• Dash: 6-layer context system. LearningMachine with agentic mode -- errors get diagnosed, fixed, and saved as learnings that are never repeated. Business rules injection.
• Codex: 2-phase memory pipeline -- extract memories from rollouts with a dedicated model, then consolidate with global locking.

Our approach (phased):

Phase 1: File-based memory (MVP)

• ~/.hermes/MEMORY.md -- curated long-term memory, agent can read/append/edit
• ~/.hermes/USER.md -- user profile the agent maintains (preferences, context, projects, communication style)
• memory tool with actions: read, append, search (simple text search within the file)
• Both files injected into system prompt (or summarized if too large)
• Agent prompted to update memory at session end or before context compression

Phase 2: Learning store (inspired by Dash)

• ~/.hermes/learnings.jsonl -- structured error patterns and discovered fixes
• save_learning + search_learnings tool actions
• Each learning: {pattern, fix, context, tags, created_at, times_used}
• Before executing risky operations, agent auto-searches learnings for known pitfalls
• Learning deduplication and consolidation over time

Phase 3: Semantic search (later)

• SQLite + sqlite-vec (or ChromaDB) for vector storage
• Embed session transcripts, memory entries, and learnings
• Hybrid search: keyword (ripgrep) + vector (embeddings) + temporal decay
• Auto-index new sessions on write
• This is the "full OpenClaw" level -- significant infrastructure

---

6. Session Transcript Search 🔍

Status: Not started Priority: Medium-High -- low-hanging fruit, very useful

Search across past session transcripts to find previous conversations, solutions, and tool outputs.

Our approach:

• CLI command: hermes sessions search <query> -- uses ripgrep over ~/.hermes/sessions/*.jsonl and logs/*.jsonl
• Agent tool: session_search(query, role_filter, limit, offset) -- same search, returns structured JSON
• Ripgrep for speed with Python fallback for environments without rg
• Filter by role (user/assistant/tool), date range, platform
• Results: session_id, line number, role, content preview centered on match
• Pagination for large result sets
• Later: integrate with Phase 3 memory (vector search over transcripts)

Files: tools/session_search_tool.py, hermes_cli/sessions.py (CLI command handler)

---

7. Local Browser Control via CDP 🌐

Status: Not started (currently Browserbase cloud only) Priority: Medium

Support local Chrome/Chromium via Chrome DevTools Protocol alongside existing Browserbase cloud backend.

What other agents do:

• OpenClaw: Full CDP-based Chrome control with snapshots, actions, uploads, profiles, file chooser, PDF save, console messages, tab management. Uses local Chrome for persistent login sessions.
• Cline: Headless browser with Computer Use (click, type, scroll, screenshot, console logs)

Our approach:

• Add a local backend option to browser_tool.py using Playwright or raw CDP
• Config toggle: browser.backend: local | browserbase | auto
• auto mode: try local first, fall back to Browserbase
• Local advantages: free, persistent login sessions, no API key needed
• Local disadvantages: no CAPTCHA solving, no stealth mode, requires Chrome installed
• Reuse the same 10-tool interface -- just swap the backend
• Later: Chrome profile management for persistent sessions across restarts

---

8. Signal Integration 📡

Status: Not started Priority: Low

New platform adapter using signal-cli daemon (JSON-RPC HTTP + SSE). Requires Java runtime and phone number registration.

Reference: OpenClaw has Signal support via signal-cli.

---

9. Plugin/Extension System 🔌

Status: Partially implemented (event hooks exist in gateway/hooks.py) Priority: Medium

Full Python plugin interface that goes beyond the current hook system.

What other agents do:

• OpenClaw: Plugin SDK with tool-send capabilities, lifecycle phase hooks (before-agent-start, after-tool-call, model-override), plugin registry with install/uninstall.
• Pi: Extensions are TypeScript modules that can register tools, commands, keyboard shortcuts, custom UI widgets, overlays, status lines, dialogs, compaction hooks, raw terminal input listeners. Extremely comprehensive.
• OpenCode: MCP client support (stdio, SSE, StreamableHTTP), OAuth auth for MCP servers. Also has Copilot/Codex plugins.
• Codex: Full MCP integration with skill dependencies.
• Cline: MCP integration + lifecycle hooks with cancellation support.

Our approach (phased):

Phase 1: Enhanced hooks

• Expand the existing gateway/hooks.py to support more events: before-tool-call, after-tool-call, before-response, context-compress, session-end
• Allow hooks to modify tool results (e.g., filter sensitive output)

Phase 2: Plugin interface

• ~/.hermes/plugins/<name>/plugin.yaml + handler.py
• Plugins can: register new tools, add CLI commands, subscribe to events, inject system prompt sections
• hermes plugin list|install|uninstall|create CLI commands
• Plugin discovery and validation on startup

Phase 3: MCP support (industry standard)

• MCP client that can connect to external MCP servers (stdio, SSE, HTTP)
• This is the big one -- Codex, Cline, and OpenCode all support MCP
• Allows Hermes to use any MCP-compatible tool server (hundreds exist)
• Config: mcp_servers list in config.yaml with connection details
• Each MCP server's tools get registered as a new toolset

---

10. Native Companion Apps 📱

Status: Not started Priority: Low

macOS (Swift/SwiftUI), iOS, Android apps connecting via WebSocket.

Prerequisite: WebSocket API on gateway (new endpoint alongside the existing HTTP flow).

What other agents do:

• OpenClaw: iOS + Android + macOS companion apps with Bonjour pairing. Voice Wake (always-on speech detection). Talk Mode (continuous conversation overlay). A2UI Canvas for agent-pushed visual content.
• OpenCode: Desktop app via Tauri (macOS, Windows, Linux). Also has a VS Code extension.

Our approach:

• MVP: Web UI with Flask/FastAPI + WebSocket (much faster to build than native apps)
• Later: Tauri desktop app wrapping the web UI
• Later: Native iOS/Android if there's demand

---

11. Evaluation System 📏

Status: Not started Priority: Medium

Systematic evaluation of agent performance for batch_runner and RL training.

What other agents do:

• Dash: Evaluation system with test cases and LLM grading

Our approach:

• LLM grader mode for batch_runner: after each trajectory, a judge model scores the result
• Action comparison: expected tool call sequences vs. actual
• String matching baselines for simple checks
• Metrics: task completion rate, tool efficiency (calls per task), cost per task
• Export to WandB for tracking across runs

---

12. Layered Context Architecture 📊

Status: Partially implemented (context files, skills, compression exist) Priority: Medium

Structured hierarchy for what goes into the system prompt, with clear priority ordering.

Current state: We have SOUL.md, AGENTS.md, skills, session context, and compression. But it's ad-hoc -- no explicit layering or budget allocation.

Our approach:

• Define explicit layers with token budgets: project context (AGENTS.md) > skills > user profile (USER.md) > learnings > memory > session context > runtime introspection
• Each layer has a max token budget; when total exceeds limit, lower-priority layers get summarized first
• Runtime introspection layer: current working directory, active processes, git status, recent file changes
• This becomes the backbone for the memory system (item 5) and subagent architecture (item 1)

---

13. Tools Wishlist 🧰

Status: Various Priority: Mixed

Diagram Rendering

• Mermaid/PlantUML text to PNG/SVG images
• Agent writes diagram code, tool renders it, returns image path
• Dependencies: mmdc (mermaid-cli) or PlantUML jar

Document Generation

• PDFs (via WeasyPrint or reportlab), Word docs (python-docx), presentations (python-pptx)
• Agent writes content, tool generates formatted document
• Templates for common formats (report, resume, letter, slides)

Canvas / Visual Workspace

• OpenClaw has this: A2UI (Agent-to-UI) -- agent pushes visual content to a Canvas surface
• For us: could be a web-based canvas (HTML/JS) that the agent can draw on
• Depends on companion app / web UI (item 10)

Coding Agent Skill

• Orchestrate Codex CLI or Claude Code via PTY mode (already supported!)
• Create a skill document that teaches the agent how to use these tools effectively
• Not a new tool -- just a well-crafted skill + PTY terminal usage

Domain Skill Packs

• Curated skill collections for specific domains
• DevOps: Docker, K8s, CI/CD, monitoring
• Data Science: pandas, scikit-learn, plotting, data cleaning
• Security: vulnerability scanning, OWASP, secrets management
• Each pack is a set of SKILL.md files installable via hermes skills install <pack>

---

NEW: Items Discovered from Agent Codebases Analysis

---

14. MCP (Model Context Protocol) Support 🔗

Status: Not started Priority: High -- this is becoming an industry standard

MCP is the protocol that Codex, Cline, and OpenCode all support for connecting to external tool servers. Supporting MCP would instantly give Hermes access to hundreds of community tool servers.

What other agents do:

• Codex: Full MCP integration with skill dependencies
• Cline: use_mcp_tool / access_mcp_resource / load_mcp_documentation tools
• OpenCode: MCP client support (stdio, SSE, StreamableHTTP transports), OAuth auth

Our approach:

• Implement an MCP client that can connect to external MCP servers
• Config: list of MCP servers in ~/.hermes/config.yaml with transport type and connection details
• Each MCP server's tools auto-registered as a dynamic toolset
• Start with stdio transport (most common), then add SSE and HTTP
• Could also be part of the Plugin system (item 9, Phase 3) since MCP is essentially a plugin protocol

---

15. Permission / Safety System 🛡️

Status: Partially implemented (dangerous command approval in gateway) Priority: Medium

Formalize the tool permission system beyond the current ad-hoc dangerous command checks.

What other agents do:

• OpenCode: Sophisticated permission system with wildcards, deny/allow/ask per tool pattern. Configurable primary_tools that bypass restrictions.
• Codex: Exec policy system for allowed/denied commands. Seatbelt (macOS) / Landlock (Linux) sandboxing. Network proxy for controlling outbound access.
• OpenClaw: Tool policy pipeline with allowlist/denylist per sandbox. Docker per-session sandboxing.

Our approach:

• Config-based permission rules: permissions.yaml with patterns like terminal.*: ask, file.write: allow, browser.*: deny
• Per-platform overrides (stricter on Discord groups than Telegram DMs)
• Tool call logging with audit trail (who ran what, when, approved by whom)
• Later: sandboxing options (Docker per-session, like OpenClaw)

---

16. Self-Learning from Errors 📖

Status: Not started Priority: Medium-High -- unique differentiator from Dash

Automatic learning loop: when tool calls fail, the agent diagnoses the error, fixes it, and saves the pattern so it never repeats the same mistake.

What Dash does:

• LearningMachine with agentic mode
• Success -> optionally save as Knowledge
• Error -> Diagnose -> Fix -> Save Learning (never repeated)
• 6 layers of context including institutional knowledge and runtime schema

Our approach:

• Part of the Memory System (item 5, Phase 2)
• ~/.hermes/learnings.jsonl stores: {pattern, error_type, fix, context, tags, created_at, times_used}
• Before executing operations that have failed before, auto-inject relevant learnings
• Agent prompted: "This is similar to a previous error. Here's what worked last time: ..."
• Consolidation: periodically merge similar learnings and increment times_used
• Could be triggered automatically on tool call errors or manually by the agent

---

17. Session Branching / Checkpoints 🌿

Status: Not started Priority: Low-Medium

Save and restore conversation state at any point. Branch off to explore alternatives without losing progress.

What other agents do:

• Pi: Full branching -- create branches from any point in conversation. Branch summary entries. Parent session tracking for tree-like session structures.
• Cline: Checkpoints -- workspace snapshots at each step with Compare/Restore UI
• OpenCode: Git-backed workspace snapshots per step, with weekly gc

Our approach:

• checkpoint tool: saves current message history + working directory state as a named snapshot
• restore tool: rolls back to a named checkpoint
• Stored in ~/.hermes/checkpoints/<session_id>/<name>.json
• For file changes: git stash or tar snapshot of working directory
• Useful for: "let me try approach A, and if it doesn't work, roll back and try B"
• Later: full branching with tree visualization

---

18. File Watcher / Project Awareness 👁️

Status: Not started Priority: Low

Monitor the working directory for changes and notify the agent of relevant updates.

What other agents do:

• Codex: File watcher for live project change detection
• OpenCode: Git info integration, worktree support

Our approach:

• Watchdog-based file watcher on the working directory
• Inject a system message when relevant files change (e.g., "Note: src/main.py was modified externally")
• Filter noise: ignore .git/, node_modules/, __pycache__/, etc.
• Useful for pair-programming scenarios where the user is also editing files
• Could also track git status changes

---

19. Heartbeat System 💓

Status: Not started Priority: Low-Medium

Periodic agent wake-up for checking reminders, monitoring tasks, and running scheduled introspection.

What other agents do:

• OpenClaw: Heartbeat runner with configurable intervals per-agent (e.g., every 30m, 10m, 15m). Ghost reminder system. Transcript pruning on heartbeat. Dynamic config update without restart. Wake-on-demand.

Our approach:

• HEARTBEAT.md file in ~/.hermes/ -- agent reads this on each heartbeat
• Configurable interval (default: 30 minutes when gateway is running)
• Runs inside an existing session (not a new one) to maintain context
• Can be used for: checking on background processes, sending reminders, running periodic tasks
• HEARTBEAT_OK response suppresses output when nothing needs attention
• Integrates with the existing cronjob system for scheduling

---

20. Programmatic Tool Calling (Code-Mediated Tool Use) 🧬

Status: Not started Priority: High -- potentially the single biggest efficiency win for agent loops

Instead of the LLM making one tool call, reading the result, deciding what to do next, making another tool call (N round trips), the LLM writes a Python script that calls multiple tools, processes results, branches on conditions, and returns a final summary -- all in one turn.

What Anthropic just shipped (Feb 2026):

Anthropic's new web_search_20260209 and web_fetch_20260209 tools use "dynamic filtering" -- Claude writes and executes Python code that calls the search/fetch tools, filters the HTML, cross-references results, retries with different queries, and returns only what's relevant. Results: +11% accuracy, -24% input tokens on average across BrowseComp and DeepsearchQA. Quora/Poe found it "achieved the highest accuracy on our internal evals" and described it as behaving "like an actual researcher, writing Python to parse, filter, and cross-reference results rather than reasoning over raw HTML in context."

Source: claude.com/blog/improved-web-search-with-dynamic-filtering

Why this matters for agent loops:

The standard agent loop is:

LLM call -> tool call -> result -> LLM call -> tool call -> result -> LLM call -> ...

Every round trip costs: a full LLM inference (prompt + generation), network latency, and the growing context window carrying all previous tool results. For a 10-step task, that's 10+ LLM calls with increasingly large contexts.

With programmatic tool calling:

LLM call -> writes Python script -> script calls tools N times, processes results,
branches on conditions, retries on failure -> returns summary -> LLM call

One LLM call replaces many. The intermediate tool results never enter the context window -- they're processed in the code sandbox and only the final summary comes back. The LLM pre-plans its decision tree in code rather than making decisions one-at-a-time in the conversation.

Which of our tools benefit most:

Tool	Current pattern (N round trips)	With programmatic calling (1 round trip)
web_search + web_extract	Search -> read results -> pick URLs -> extract each -> read each -> synthesize	Script: search, fetch top 5, extract relevant sections, cross-reference, return summary
browser (10 tools)	navigate -> snapshot -> click -> snapshot -> type -> snapshot -> ...	Script: navigate, loop through elements, extract data, handle pagination, return structured result
file ops (read, search, patch)	Search for pattern -> read matching files -> decide which to patch -> patch each	Script: search, read all matches, filter by criteria, apply patches, verify, return diff summary
session_search	Search -> read results -> search again with refined query -> ...	Script: search with multiple queries, deduplicate, rank by relevance, return top N
terminal	Run command -> check output -> run follow-up -> check again -> ...	Script: run command, parse output, branch on exit code, run follow-ups, return final state

The hard problem: where does the code run?

Our tools don't all live in one place. The terminal backend can be local, Docker, Singularity, SSH, or Modal. Browser runs on Browserbase cloud. Web search/extract are Firecrawl API calls. File ops go through whatever terminal backend is active. Vision, image gen, TTS are all remote APIs.

If we just "run Python in the terminal," we hit a wall:

• Docker/Modal/SSH backends: The remote environment doesn't have our Python tool code, our API keys, our handle_function_call dispatcher, or any of the hermes-agent packages. It's a bare sandbox.
• Local backend: Could import our code directly, but that couples execution to local-only and creates a security mess (LLM-generated code running in the agent process).
• API-based tools (web, browser, vision): These need API keys and specific client libraries that aren't in the terminal backend.

The code sandbox is NOT the terminal backend. This is the key insight. The sandbox runs on the agent host machine (where run_agent.py lives), separate from both the LLM and the terminal backend. It calls tools through the same handle_function_call dispatcher that the normal agent loop uses. No inbound network connections needed -- everything is local IPC on the agent host.

Architecture: Local subprocess with Unix domain socket RPC

Agent Host (where run_agent.py runs)

  ┌──────────────┐         ┌──────────────────────┐
  │ Code Sandbox │  UDS    │ Agent Process         │
  │ (child proc) │ ◄─────► │ (parent)              │
  │              │ (local) │                       │
  │ print()──────┼─stdout──┼─► captured as output  │
  │ errors───────┼─stderr──┼─► captured for debug  │
  │              │         │                       │
  │ hermes_tools │         │ handle_function_call  │
  │ .web_search()├─socket──┼─► web_search ─────────┼──► Firecrawl API
  │ .terminal()  ├─socket──┼─► terminal ───────────┼──► Docker/SSH/Modal
  │ .browser_*() ├─socket──┼─► browser ────────────┼──► Browserbase
  │ .read_file() ├─socket──┼─► read_file ──────────┼──► terminal backend
  └──────────────┘         └──────────────────────┘

The sandbox is a child process on the same machine. RPC goes over a Unix domain socket (not stdin/stdout/stderr -- those stay free for their natural purposes). The parent dispatches each tool call through the existing handle_function_call -- the exact same codepath the normal agent loop uses. Works with every terminal backend because the sandbox doesn't touch the terminal backend directly.

RPC transport: Unix domain socket

Why not stdin/stdout/stderr for RPC?

• stdout is the script's natural output channel (print()). Multiplexing RPC and output on the same stream requires fragile marker parsing. Keep it clean: stdout = final output for the LLM.
• stderr is for Python errors, tracebacks, warnings, and logging output. Multiplexing RPC here means any stray logging.warning() or exception traceback corrupts the RPC stream.
• Extra file descriptors (fd 3/4) work on Linux/macOS but are clunky with subprocess.Popen.

A Unix domain socket gives a clean dedicated RPC channel:

1. Parent creates a temp UDS: /tmp/hermes_rpc_<uuid>.sock
2. Parent starts listening (single-client, since there's one sandbox)
3. Parent spawns child with HERMES_RPC_SOCKET=/tmp/hermes_rpc_<uuid>.sock in env
4. Child's hermes_tools module connects to the socket on first tool call
5. Protocol: newline-delimited JSON. Child writes {"tool": "web_search", "args": {...}}\n, reads {"result": ...}\n back
6. Parent reads each request, calls handle_function_call, writes the response
7. After child exits, parent cleans up the socket file

Channels stay clean:

• stdout → captured by parent as the tool's return value to the LLM
• stderr → captured by parent for error reporting (included in response on failure)
• UDS → dedicated tool call RPC (invisible to the script's normal I/O)

Works on Linux and macOS (our target platforms). Windows fallback: named pipes or the marker-on-stderr approach if we ever need it.

The auto-generated hermes_tools module

The parent writes this into a temp directory before spawning the child. Each function is a thin RPC stub:

# Auto-generated: /tmp/hermes_sandbox_<uuid>/hermes_tools.py
import json, os, socket

_sock = None

def _connect():
    global _sock
    if _sock is None:
        _sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
        _sock.connect(os.environ["HERMES_RPC_SOCKET"])
        _sock.settimeout(300)  # 5 min max per tool call
    return _sock

def _call(tool_name, args):
    """RPC: send tool call to parent, get result back."""
    conn = _connect()
    request = json.dumps({"tool": tool_name, "args": args}) + "\n"
    conn.sendall(request.encode())
    # Read response (newline-delimited)
    chunks = []
    while True:
        data = conn.recv(65536)
        if not data:
            raise RuntimeError("Agent process disconnected")
        chunks.append(data.decode())
        if chunks[-1].endswith("\n"):
            break
    raw = "".join(chunks).strip()
    # Tool responses are JSON strings; parse them into dicts
    result = json.loads(raw)
    if isinstance(result, str):
        try:
            return json.loads(result)
        except (json.JSONDecodeError, TypeError):
            return result
    return result

# --- Tool functions (one per enabled tool) ---

def web_search(query):
    """Search the web. Returns dict with 'results' list."""
    return _call("web_search", {"query": query})

def web_extract(urls):
    """Extract content from URLs. Returns markdown text."""
    return _call("web_extract", {"urls": urls})

def read_file(path, offset=1, limit=500):
    """Read a file. Returns dict with content and metadata."""
    return _call("read_file", {"path": path, "offset": offset, "limit": limit})

def terminal(command, timeout=None):
    """Run a shell command. Returns dict with stdout, exit_code."""
    return _call("terminal", {"command": command, "timeout": timeout})

def search(pattern, target="content", path=".", file_glob=None, limit=50):
    """Search file contents or find files."""
    return _call("search", {"pattern": pattern, "target": target,
                             "path": path, "file_glob": file_glob, "limit": limit})

# ... generated for each enabled tool in the session

This module is generated dynamically because the available tools vary per session and per toolset configuration. The generator reads the session's enabled tools and emits a function for each one that's on the sandbox allow-list.

What Python libraries are available in the sandbox

The sandbox runs the same Python interpreter as the agent. Available imports:

Python standard library (always available): json, re, math, csv, datetime, collections, itertools, textwrap, difflib, html, urllib.parse, pathlib, hashlib, base64, string, functools, operator, statistics, io, os.path

Not restricted but discouraged via tool description: subprocess, socket, requests, urllib.request, os.system -- the tool description says "use hermes_tools for all I/O." We don't hard-block these because the user already trusts the agent with terminal(), which is unrestricted shell access. Soft-guiding the LLM via the description is sufficient. If it occasionally uses import os; os.listdir() instead of hermes_tools.search(), no real harm done.

The tool description tells the LLM:

Available imports:
- from hermes_tools import web_search, web_extract, read_file, terminal, ...
- Python standard library: json, re, math, csv, datetime, collections, etc.

Use hermes_tools for all I/O (web, files, commands, browser).
Use stdlib for processing between tool calls (parsing, filtering, formatting).
Print your final result to stdout.

Platform support

Linux / macOS: Fully supported. Unix domain sockets work natively.

Windows: Not supported. AF_UNIX nominally exists on Windows 10 17063+ but is unreliable in practice, and Hermes-Agent's primary target is Linux/macOS (bash-based install, systemd gateway, etc.). The execute_code tool is disabled at startup on Windows:

import sys
SANDBOX_AVAILABLE = sys.platform != "win32"

def check_sandbox_requirements():
    return SANDBOX_AVAILABLE

If the LLM tries to use execute_code on Windows, it gets: {"error": "execute_code is not available on Windows. Use normal tool calls instead."}. The tool is excluded from the tool schema entirely on Windows so the LLM never sees it.

Which tools to expose in the sandbox (full audit)

The purpose of the sandbox is reading, filtering, and processing data across multiple tool calls in code, collapsing what would be many LLM round trips into one. Every tool needs to justify its inclusion against that purpose. The parent only generates RPC stubs for tools that pass this filter AND are enabled in the session.

Every tool, one by one:

Tool	In sandbox?	Reasoning
web_search	YES	Core use case. Multi-query, cross-reference, filter.
web_extract	YES	Core use case. Fetch N pages, parse, keep only relevant sections.
read_file	YES	Core use case. Bulk read + filter. Note: reads files on the terminal backend (Docker/SSH/Modal), not the agent host -- this is correct and intentional.
search	YES	Core use case. Find files/content, then process matching results.
terminal	YES (restricted)	Command chains with branching on exit codes. Foreground only -- background, check_interval, and pty parameters are stripped/blocked.
write_file	YES (with caution)	Scripts need to write computed outputs (generated configs, processed data). Partial-write risk if script fails midway, but same risk as normal tool calls.
patch	YES (with caution)	Bulk search-and-replace across files. Powerful but risky if the script's patch logic has bugs. The upside: script can read → patch → verify in a loop, which is actually safer than blind patching.
browser_navigate	YES	Browser automation loops are one of the biggest wins.
browser_snapshot	YES	Needed for reading page state in browser loops. Parent passes user_task from the session context.
browser_click	YES	Core browser automation.
browser_type	YES	Core browser automation.
browser_scroll	YES	Core browser automation.
browser_back	YES	Navigation within browser loops.
browser_press	YES	Keyboard interaction in browser loops.
browser_close	NO	Ends the entire browser session. If the script errors out after closing, the agent has no browser to recover with. Too destructive for unsupervised code.
browser_get_images	NO	Niche. Usually paired with vision analysis, which is excluded.
browser_vision	NO	This calls the vision LLM API -- expensive per call and requires LLM reasoning on the result. Defeats the purpose of avoiding LLM round trips.
vision_analyze	NO	Expensive API call per invocation. The LLM needs to SEE and reason about images directly, not filter them in code. One-shot nature.
mixture_of_agents	NO	This IS multiple LLM calls. Defeats the entire purpose.
image_generate	NO	Media generation. One-shot, no filtering logic benefits.
text_to_speech	NO	Media generation. One-shot.
process	NO	Background process management from an ephemeral script is incoherent. The script exits, but the process lives on -- who monitors it?
skills_list	NO	Skills are knowledge for the LLM to read and reason about. Loading a skill inside code that can't reason about it is pointless.
skill_view	NO	Same as above.
schedule_cronjob	NO	Side effect. Should not happen silently inside a script.
list_cronjobs	NO	Read-only but not useful in a code-mediation context.
remove_cronjob	NO	Side effect.
send_message	NO	Cross-platform side effect. Must not happen unsupervised.
todo_write	NO	Agent-level conversational state. Meaningless from code.
todo_read	NO	Same.
clarify	NO	Requires interactive user input. Can't block in a script.
execute_code	NO	No recursive sandboxing.
All RL tools	NO	Separate domain with its own execution model.

Summary: 14 tools in, 28+ tools out. The sandbox exposes: web_search, web_extract, read_file, write_file, search, patch, terminal (restricted), browser_navigate, browser_snapshot, browser_click, browser_type, browser_scroll, browser_back, browser_press.

The allow-list is a constant in code_execution_tool.py, not derived from the session's enabled toolsets. Even if the session has vision_analyze enabled, it won't appear in the sandbox. The intersection of the allow-list and the session's enabled tools determines what's generated.

Error handling

Scenario	What happens
Syntax error in script	Child exits immediately, traceback on stderr. Parent returns stderr as the tool response so the LLM sees the error and can retry.
Runtime exception	Same -- traceback on stderr, parent returns it.
Tool call fails	RPC returns the error JSON (same as normal tool errors). Script decides: retry, skip, or raise.
Unknown tool called	RPC returns {"error": "Unknown tool: foo. Available: web_search, read_file, ..."}.
Script hangs / infinite loop	Killed by timeout (SIGTERM, then SIGKILL after 5s). Parent returns timeout error.
Parent crashes mid-execution	Child's socket connect/read fails, gets a RuntimeError, exits.
Child crashes mid-execution	Parent detects child exit via process.poll(). Collects partial stdout + stderr.
Slow tool call (e.g., terminal make)	Overall timeout covers total execution. One slow call is fine if total is under limit.
Tool response too large in memory	web_extract can return 500KB per page. If the script fetches 10 pages, that's 5MB in the child's memory. Not a problem on modern machines, and the whole point is the script FILTERS this down before printing.
User interrupt (new message on gateway)	Parent catches the interrupt event (same as existing _interrupt_event in terminal_tool), sends SIGTERM to child, returns {"status": "interrupted"}.
Script tries to call excluded tool	RPC returns {"error": "Tool 'vision_analyze' is not available in execute_code. Use it as a normal tool call instead."}
Script calls terminal with background=True	RPC strips the parameter and runs foreground, or returns an error. Background processes from ephemeral scripts are not supported.

The tool response includes structured metadata:

{
  "status": "success | error | timeout | interrupted",
  "output": "...",
  "errors": "...",
  "tool_calls_made": 7,
  "duration_seconds": 12.3
}

Resource limits

• Timeout: 120 seconds default (configurable via config.yaml). Parent sends SIGTERM, waits 5s, SIGKILL.
• Tool call limit: max 50 RPC tool calls per execution. After 50, further calls return an error. Prevents infinite tool-call loops.
• Output size: stdout capped at 50KB. Truncated with [output truncated at 50KB]. Prevents the script from flooding the LLM's context with a huge result (which would defeat the purpose).
• Stderr capture: capped at 10KB for error reporting.
• No recursive sandboxing: execute_code is not in the sandbox's tool list.
• Interrupt support: respects the same _interrupt_event mechanism as terminal_tool. If the user sends a new message while the sandbox is running, the child is killed and the agent can process the interrupt.

Tool call logging and observability

Each tool call made inside the sandbox is logged to the session transcript for debugging, but NOT added to the LLM conversation history (that's the whole point -- keeping intermediate results out of context).

The parent logs each RPC-dispatched call:

{"type": "sandbox_tool_call", "tool": "web_search", "args": {"query": "..."}, "duration": 1.2}
{"type": "sandbox_tool_call", "tool": "web_extract", "args": {"urls": [...]}, "duration": 3.4}

These appear in the JSONL transcript and in verbose logging, but the LLM only sees the final execute_code response containing the script's stdout.

For the gateway (messaging platforms): show one typing indicator + notification for the entire execute_code duration. Internal tool calls are silent. Later enhancement: progress updates like "execute_code (3/7 tool calls)".

Stateful tools work correctly

Tools like terminal (working directory, env vars) and browser_* (page state, cookies) maintain state per task_id. The parent passes the session's task_id to every RPC-dispatched handle_function_call. So if the script runs:

terminal("cd /tmp")
terminal("pwd")  # returns /tmp -- state persists between calls

This works because both calls go through the same terminal environment, same as normal tool calls.

Each execute_code invocation is stateless

The sandbox subprocess is fresh each time. No Python state carries over between execute_code calls. If the agent needs state across multiple execute_code calls, it should:

• Output the state as part of the result, then pass it back in the next script as a variable
• Or use the tools themselves for persistence (write to a file, then read it in the next script)

The underlying tools are stateful (same terminal session, same browser session), but the Python sandbox is not.

When should the LLM use execute_code vs normal tool calls?

This goes in the tool description:

Use execute_code when:

• You need 3+ tool calls with processing logic between them
• You need to filter/reduce large tool outputs before they enter your context
• You need conditional branching (if X then do Y, else do Z)
• You need to loop (fetch N pages, process N files, retry on failure)

Use normal tool calls when:

• Single tool call with no processing needed
• You need to see the full result and apply complex reasoning (the LLM is better at reasoning than the code it writes)
• The task requires human interaction (clarify tool)

Open questions for implementation

1. Should the parent block its main thread while the sandbox runs? Currently handle_function_call is synchronous, so yes -- same as any other tool call. For long sandbox runs (up to 120s), the gateway's typing indicator stays active. The agent can't process new messages during this time, but it can't during any long tool call either. Interrupt support (above) handles the "user sends a new message" case.

2. Should browser_snapshot pass user_task? handle_browser_function_call accepts user_task for task-aware content extraction. The parent should pass the user's original query from the session context when dispatching sandbox browser calls.

3. Terminal parameter restrictions: The sandbox version of terminal should strip/ignore: background=True (no background processes from ephemeral scripts), check_interval (gateway-only feature for background watchers), pty=True (interactive PTY makes no sense in a script). Only command, timeout, and workdir are passed through.

Future enhancements (not MVP)

• Concurrent tool calls via threading: The script could use ThreadPoolExecutor to fetch 5 URLs in parallel. Requires the UDS client to be thread-safe (add a threading.Lock around socket send/receive). Significant speedup for I/O-bound workflows like multi-page web extraction.
• Streaming progress to gateway: Instead of one notification for the entire run, send periodic progress updates ("execute_code: 3/7 tool calls, 12s elapsed").
• Persistent sandbox sessions: Keep the subprocess alive between execute_code calls so Python variables carry over. Adds complexity but enables iterative multi-step workflows where the agent refines its script.
• RL/batch integration: Atropos RL environments use ToolContext instead of handle_function_call. Would need an adapter so the RPC bridge dispatches through the right mechanism per execution context.
• Windows support: If there's demand, fall back to TCP localhost (127.0.0.1:random_port) instead of UDS. Same protocol, different transport. Security concern: localhost port is accessible to other processes on the machine. Could mitigate with a random auth token in the RPC handshake.

Relationship to other items

• Subagent Architecture (#1): A code sandbox that calls tools IS a lightweight subagent without its own LLM inference. This handles many of the "mechanical multi-step" cases (search+filter, bulk file ops, browser loops) at near-zero LLM cost. Full subagents are still needed for tasks requiring LLM reasoning at each step.
• Browser automation (#7): Biggest win. Browser workflows are 10+ round trips today. A script that navigates, clicks, extracts, paginates in a loop collapses that to 1 LLM turn.
• Web search: Directly matches Anthropic's dynamic filtering results.
• File ops: Bulk read-search-patch workflows become one call.

Files: tools/code_execution_tool.py (subprocess management, UDS server, RPC dispatch, hermes_tools generator), tool schema in model_tools.py

---

Implementation Priority Order

Tier 1 (High impact, foundation for everything else):

1. Programmatic Tool Calling (code-mediated tool use) -- #20
2. Memory System (Phase 1: MEMORY.md + USER.md) -- #5
3. Planning & Task Management (todo tool) -- #2
4. Session Transcript Search -- #6
5. Self-Learning from Errors -- #16

Tier 2 (High impact, more complex): 6. Subagent Architecture -- #1 (partially solved by #20) 7. MCP Support -- #14 8. Interactive Clarifying Questions -- #4 9. Dynamic Skills Expansion (create/edit/delete) -- #3

Tier 3 (Quality of life, polish): 10. Permission / Safety System -- #15 11. Local Browser Control via CDP -- #7 12. Layered Context Architecture -- #12 13. Plugin/Extension System (enhanced hooks first) -- #9 14. Evaluation System -- #11

Tier 4 (Nice to have, longer term): 15. Heartbeat System -- #19 16. Session Branching / Checkpoints -- #17 17. File Watcher -- #18 18. Signal Integration -- #8 19. Tools Wishlist items -- #13 20. Native Companion Apps -- #10