Research

Research is the execution engine of MAARS. It plans concrete work, runs task-level agent sessions, verifies outputs, and decides whether another iteration is warranted.

01What Research is responsible for

02Outer loop

Calibrate runs once; the remaining phases repeat until evaluation is satisfied or the iteration limit is reached.

03Task model

Research breaks the problem into atomic tasks arranged as a dependency-aware DAG.

04Task execution cycle

Each task is a bounded mini-session: execute, verify, maybe retry, maybe redecompose.

05Iteration policy

Research is allowed to loop, but it is biased toward stopping rather than exploring forever.

06Runtime guards

Execution is wrapped in deterministic runtime constraints so open-ended agent behavior stays bounded.

07Outputs consumed by downstream stages

Research writes both raw outputs and normalized summaries so Write can stay grounded.

08Code map

The Research implementation spans one main stage file plus smaller decomposition and summary helpers.