The rule that an instruction can only change if it cites a line in agent-misses.md or a measurement is the part I would copy. Instruction files tend to grow every time someone reads a new tip, and nobody removes anything. A one-page limit with add one, remove one keeps it honest.
The numbers make the case well. 98% cache reads says the agents are mostly re-reading long context, and nearly a quarter of tool calls being searches suggests they are rediscovering facts that were already written down. Mapping short requests to a fixed procedure, design doc, approval, implement, then review by a different agent, should cut both.
At Ascendra Ventures we work on client systems with similar limits, read-only repos and no direct database, and the separation between the agent that builds and the one that verifies has been the most reliable habit we have. A mapping table where every row carries its evidence file is a nice touch for legacy migration.
For the Nov 25 results, will you also track how many "done" reports the verifier rejected?
I would track those rejected done reports. I would also count the times an agent declared done before the verification step in the procedure had run, because those are process misses even when the verifier later catches them.
Ascendra Ventures
The rule that an instruction can only change if it cites a line in agent-misses.md or a measurement is the part I would copy. Instruction files tend to grow every time someone reads a new tip, and nobody removes anything. A one-page limit with add one, remove one keeps it honest.
The numbers make the case well. 98% cache reads says the agents are mostly re-reading long context, and nearly a quarter of tool calls being searches suggests they are rediscovering facts that were already written down. Mapping short requests to a fixed procedure, design doc, approval, implement, then review by a different agent, should cut both.
At Ascendra Ventures we work on client systems with similar limits, read-only repos and no direct database, and the separation between the agent that builds and the one that verifies has been the most reliable habit we have. A mapping table where every row carries its evidence file is a nice touch for legacy migration.
For the Nov 25 results, will you also track how many "done" reports the verifier rejected?