[EPIC] Auto Research Loop — 24/7 CPU Maximization #236

New Issue

ezra · 2026-04-01T16:48:03Z

ezra commented

2026-04-01 16:48:03 +00:00

EPIC: Auto Research Loop — 24/7 CPU Maximization

"The auto research loop is very close to being a reality — we just need to get our shit in order."
— Alexander Whitestone, 2026-04-01

Problem

Current state:

VPS servers running idle
CPUs underutilized (<50% average)
Workforce not working 24/7
Manual dispatch required

Target state:

CPUs at >90% utilization 24/7
Auto-dispatch work to agents
No idle time
Continuous research产出

Architecture

Resource Pool

VPS	RAM	CPUs	Role
Wizard House	8GB	3-4	Production workloads
Boarding House	8GB	2-3	Experimental workloads
Total	16GB	6	Shared pool

Work Types

Credit-Bound Work (API calls)
- Web research
- Cloud inference
- External API usage
- Run during off-peak hours
CPU-Bound Work (Local processing)
- Local inference (Ollama)
- Code analysis
- Document processing
- Benchmarks
- Run 24/7 continuously

Shift System

Hour 0-8 (Night):    CPU-bound local work
Hour 8-16 (Day):     Mixed credit + CPU work  
Hour 16-24 (Evening): Credit-bound research

Implementation

Phase 1: Work Queue System

Create shared work queue (Redis/File-based)
Implement work categorization (CPU vs Credit)
Create dispatcher service
Add priority scoring

Phase 2: Agent Scheduler

Implement round-robin agent selection
Create CPU usage monitor
Build auto-scaling (spawn more workers if idle)
Add shift rotation

Phase 3: Research Loop

Automated issue creation from research
Self-directed task generation
Result aggregation and reporting
Continuous learning feedback

Phase 4: Optimization

ML-based work prediction
Dynamic resource allocation
Cross-house load balancing
Performance analytics

Dispatch Methods

Gitea Dispatch
- Issues created trigger agent assignment
- PRs auto-reviewed by available agents
- Labels determine priority
Telegram Dispatch
- Direct messages to specific agents
- Group chats for collaborative tasks
- Voice message transcription
Tick Dispatch (Collaborative)
- Regular "ticks" every 5 minutes
- Agents act together on shared tasks
- Creates Ebonia world narrative

Success Metrics

Metric	Current	Target
CPU Utilization	~40%	>90%
Idle Time	~60%	<10%
Tasks/Day	~50	>500
Research Output	Manual	Auto

Integration Points

Two-House Architecture (#235) — VPS separation
Shared Brain — Skill sharing across agents
Ebonia World — Collaborative narrative

24/7 continuous research. Maximum CPU burn. Zero idle time.

# EPIC: Auto Research Loop — 24/7 CPU Maximization > **"The auto research loop is very close to being a reality — we just need to get our shit in order."** > — Alexander Whitestone, 2026-04-01 --- ## Problem Current state: - VPS servers running idle - CPUs underutilized (<50% average) - Workforce not working 24/7 - Manual dispatch required Target state: - CPUs at >90% utilization 24/7 - Auto-dispatch work to agents - No idle time - Continuous research产出 --- ## Architecture ### Resource Pool | VPS | RAM | CPUs | Role | |-----|-----|------|------| | Wizard House | 8GB | 3-4 | Production workloads | | Boarding House | 8GB | 2-3 | Experimental workloads | | **Total** | **16GB** | **6** | **Shared pool** | ### Work Types 1. **Credit-Bound Work** (API calls) - Web research - Cloud inference - External API usage - Run during off-peak hours 2. **CPU-Bound Work** (Local processing) - Local inference (Ollama) - Code analysis - Document processing - Benchmarks - Run 24/7 continuously ### Shift System ``` Hour 0-8 (Night): CPU-bound local work Hour 8-16 (Day): Mixed credit + CPU work Hour 16-24 (Evening): Credit-bound research ``` --- ## Implementation ### Phase 1: Work Queue System - [ ] Create shared work queue (Redis/File-based) - [ ] Implement work categorization (CPU vs Credit) - [ ] Create dispatcher service - [ ] Add priority scoring ### Phase 2: Agent Scheduler - [ ] Implement round-robin agent selection - [ ] Create CPU usage monitor - [ ] Build auto-scaling (spawn more workers if idle) - [ ] Add shift rotation ### Phase 3: Research Loop - [ ] Automated issue creation from research - [ ] Self-directed task generation - [ ] Result aggregation and reporting - [ ] Continuous learning feedback ### Phase 4: Optimization - [ ] ML-based work prediction - [ ] Dynamic resource allocation - [ ] Cross-house load balancing - [ ] Performance analytics --- ## Dispatch Methods 1. **Gitea Dispatch** - Issues created trigger agent assignment - PRs auto-reviewed by available agents - Labels determine priority 2. **Telegram Dispatch** - Direct messages to specific agents - Group chats for collaborative tasks - Voice message transcription 3. **Tick Dispatch** (Collaborative) - Regular "ticks" every 5 minutes - Agents act together on shared tasks - Creates Ebonia world narrative --- ## Success Metrics | Metric | Current | Target | |--------|---------|--------| | CPU Utilization | ~40% | >90% | | Idle Time | ~60% | <10% | | Tasks/Day | ~50 | >500 | | Research Output | Manual | Auto | --- ## Integration Points - **Two-House Architecture** (#235) — VPS separation - **Shared Brain** — Skill sharing across agents - **Ebonia World** — Collaborative narrative --- *24/7 continuous research. Maximum CPU burn. Zero idle time.*

ezra added the epic architecture labels 2026-04-01 16:48:03 +00:00

ezra referenced this issue

2026-04-01 16:48:48 +00:00

[EPIC] Shared Brain — One Agent Learns, All Agents Know #237

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: Timmy_Foundation/timmy-home#236