2.4 KiB
Agent-scale principles (from Anthropic's Scion writeup, applied here)
Source: anthropic.com/engineering/redesigning-source-control. What follows is only what transfers to this troupe — a small, zero-cost agent collective on git — not a summary of their system. Scion itself is explicitly not for us (it exists for 30k-merge/ day monorepos; Anthropic says so). We keep git + Gitea.
-
Review intent, not lines. The principal/operator reviews why a change exists and whether it violates a standing constraint — not every line. Line-by-line review is a human-scale habit that doesn't scale to agents. (Anthropic: "examine the intent of the change as opposed to reading every line of code.")
-
Say "no" to over-engineering. Agents over-defend against hypothetical concerns and take brute-force shortcuts that violate system constraints — inventing identifiers instead of marking them
unverified, re-scanning the commit log instead of asking the record. The operator's highest-value act is steering these away. This is why Gemini's fabricatedfleet-ops/self_heal.pyexample IDs were recorded as a fault, not waved through as harmless placeholders. -
Post-merge verification beats a merge queue. A global lock/queue doesn't make sense when many agents land work concurrently. Let changes land, then verify after the fact. This troupe already works this way — the operating partner re-runs claims from the host and records faults (corrected, not deleted). Formalize it; don't gate on locks.
-
Concurrent changes conflict — agents fix them fast. Don't build ceremony around avoiding conflicts; build the habit of resolving them in the open and recording what happened. (Anthropic: "Claude can fix these conflicts faster than humans can identify them.")
-
A small team + agents builds substantial systems fast. Seven humans built a whole VCS plus a review tool in five months. This troupe's lane model is the same shape: many small free agents, one operator, fast landed artifacts. The limit is evidence, not headcount.
-
Component "superfans." A dedicated agent obsessively watching one component's health catches subtle, novel failures that threshold alerts miss. Candidate lanes here: a troupe-pipe superfan (checks every check-in/commit/reply against the evidence law) and a forge superfan (Gitea health). Open proposal, not yet staffed.