Change the crew.
Rebalance the promises.
Build a variation from the edited prompt and compare it with the supplied program. The expected effects below are predictions for that build; testing a fresh agent-generated version is your next experiment.
A one-source mini relay
The planner chooses 1-3 unique existing source IDs / Use at most five adapter calls per run
Choose exactly one approved source and use at most three adapter calls: one plan, one worker and one review.
A successful run records one evidence note and reaches review in three calls. Its coverage centers on one selected source.
The budget follows the work graph: 1+1+1 replaces 1+3+1. Reducing the cap alone could stop a three-worker plan before its review.
Use one fictional note; verify three calls, an exact quote and approval before export. Then try an invalid quote and a restart.
Change planner validation, stored maxCalls, the actual budget guard, checkpoint validation and the visible counter together.
Proposed variant: a fresh agent-generated build and test remain the next steps.
Parallel workers, the same total budget
one plan, up to three workers, one review
After the plan, run selected workers concurrently; reserve every call atomically, save results by source ID, and review only after all workers finish.
Worker replies may arrive in any order. The UI can show separate progress. Measure elapsed time in your new build to find out whether concurrency helps.
Concurrent calls change scheduling and shared-state coordination. Each call still consumes budget, and review waits for the worker results.
Use fake workers with different delays. Require exactly one call and result per selected source, a budget cap and a review that starts after all results arrive.
The supplied program is sequential and locks an entire step. Ask the agent to implement a new coordinator that reserves calls atomically, collects worker results and starts review after completion.
Proposed variant: a fresh agent-generated build and test remain the next steps.