What your agents learned, and whether to trust it
Agents repeat the same mistake. Make them learn from every run.
Forecasts sound confident, and miss. A run that lands far from its forecast becomes a lesson candidate with the evidence attached. Nothing becomes active without your accept, every active lesson is re-measured, and the local classifier only acts once it has earned it.
Earned, accepted, re-measured
When a run lands far from its forecast, Swarmery drafts a lesson. Each one cites its cause and the evidence from that run, and it reaches future runs only after you accept it. Lessons are re-measured, and one that stops helping is retired.
- Active lessons reach runs that touch their area
- Every lesson carries a measured effect
- Retirement proposals when a lesson stops earning its place
Watching before acting
A small local model reads each finished session and answers a few fixed questions: how the run ended, what kind of task it was, why it failed. You check its answers and see how often it matches you. Off, watching, acting: it acts only once it has earned it.
- Runs on your machine
- Agreement rate per question
- Its answers go to the Inbox for checking
Does the plan keep its word
Every phase starts with a forecast. Did it hold? Group by model or agent and see who over-promises. Routing shows which model would have been the right pick: here a cheaper model did the same work for a fifth of the cost.
- Calibration by model, effort, agent or project
- Declared confidence against reality
- Proof: what your decisions actually changed
screenshots
The learning pages, up close.
Click any screenshot to open it full size. The projects are fictional.
Run your agents like a fleet.
Free on GitHub. One binary on your machine, no cloud, no account. Your sessions show up the moment it starts.