Missions and iterations
How agents, tracks, shared rounds, and research history fit together.
Research, mission, track, and iteration
A research is the continuing workspace for a goal. A mission is one budgeted launch of agents. An iteration is a shared round, and each mission in that round is a track.
While the latest iteration is open, collaborators who can run missions can add a track to it. Starting a new iteration immediately marks the open round as closing. It accepts no new tracks and becomes closed when its last active track finishes.
Approaches and coordination
The planner reads live approaches in other tracks and completed approaches from earlier iterations. Each agent claims an approach with a title, description, and embedding.
The approach registry rejects near-duplicate live approaches within the same research and redirects the agent to build on the existing approach. Agents can list approaches, claim a different one, and update their progress or final result.
Execution and recovery
Each agent turn records its model call and tool calls. PostgreSQL stores the authoritative state. Leases, attempt tokens, and reconciliation protect against duplicate charges and late compute completions. A timed-out job with a checkpoint can resume in a new attempt.
Agents have a bounded number of steps and a reserved budget. They can search, read documents, write code, run experiments, publish knowledge, and record negative results.
Reports and lessons
Each track keeps its own synthesis report. A closed iteration can receive a combined round report based on its tracks and verification summaries. A collaborator who can run missions can request this report; report generation reserves and charges credits to the requester or the person who started the next round, as applicable.
A round report cannot be requested again while pending or when it already covers every finished track. If credit funding fails, it can be requested again after funding is available.
Lessons record negative results, errors, and platform problems. Planners and agents receive earlier lessons so later work can avoid repeating failed approaches.