Runtime + autoresearch · open
Where the loop closes around the agent.
The runtime executes agent .ts files. Autoresearch tunes them against the Reality Twin — every run scored, every score feeding the next iteration.
What runs here
The orchestrator
Turns a compiled plan into a running execution pipeline. The runtime sees only compiled .ts — never DOIL itself.
4-layer evaluation
Each agent passes four evaluation layers before certification — vendor-neutral, simulator-backed.
Autoresearch loop
Reads the compiler's constraint comments. Tunes the agent .ts directly against simulator outputs.
Evaluation corpus
Every run appends a scored trace. A clustered visual map shows which agents are tuned hardest.
The loop
factory → DOIL → runtime ⇄ simulators → autoresearch → factory