| name | dont-collapse-the-architecture |
| description | When tempted to collapse an ambitious-but-unproven architecture (multi-agent topology, context-lifecycle management, a recursive loop system) into a dumb/old pattern because an early A/B looked marginal — don't. Marginal-early almost always means the regime that makes the architecture pay off wasn't active. Find that regime, build the missing competency, then judge. |
Don't collapse the architecture
The reductive instinct — "it's unproven, an A/B was marginal, simplify it to the dumb/old pattern" — deletes the exact thing that would have worked once the missing competency was built. Marginal-early ≠ worthless. It almost always means the constraint that makes the architecture pay off was not active in the test.
This is a one-way door: collapsing throws away the thing whose value was latent. Default to holding the architecture and finding its regime, not collapsing it.
The worked case: agents are a context-management problem
Agent capability is, at bottom, a context-management and knowledge problem. The finite context window is the binding constraint; quality rots as it fills. The value of a multi-agent topology is latent — unlocked by one competency: expert management of the context lifecycle (close a chapter → checkpoint to external tracking state; open a chapter → respawn a fresh agent against that state; recurse).
A naive "does 2 agents beat 1" A/B on a task that fits in one context window will tie — because the context constraint never bit. That tie says nothing about the topology; it says you measured in the wrong regime. (See docs/research/smart-loops-context-lifecycle.md in agent-runtime, and the Autodata null that was an extractive task + a memorized doc, not "the loop doesn't work.")
The rule — when you catch yourself about to simplify on marginal evidence
- Name the regime where the architecture is supposed to pay off. Was the test in that regime? (For context-lifecycle work: long-horizon tasks where context exhaustion is the binding constraint.) If not, the result is uninformative — do not act on it.
- Name the competency that's missing. Marginal-without-the-competency ≠ marginal-with it. (Here: the chapter close/open policy.)
- Build the competency and test in the right regime before reducing. Only a result in the regime, with the competency built can justify collapsing.
- If you still feel the pull to simplify, say out loud: "I'm about to delete something whose value is latent." Then don't — find the regime instead.
Hold the thesis. Find where it bites. Build the policy. Then — and only then — judge.
Then consider (post-hook)
- About to A/B the architecture? →
calibrate-before-measure: prove the metric discriminates a good policy from a bad one AND that the test sits in the regime where the constraint is active. A tie outside that regime is uninformative, not a verdict.
- Sitting on a marginal/null result? →
push-past-easy: the next move is the experiment in the right regime with the competency built — never a collapse to the old pattern.