Recursive
Reason-act-observe loop converging toward a goal.
Verified against arxiv.org
A recursive agent loops on itself until a stopping condition is met: reason, act, observe the result, then feed that observation into the next pass. Rather than a fixed pipeline it re-plans each turn with fresh evidence. This is the shape behind most coding and research agents: the model proposes a step, a tool executes it, the outcome revises the plan, and the loop repeats until the task is done or a budget is hit. In anatomy terms it is the reasoning core wired to its own working memory and effectors in a tight feedback cycle.
The reasoning core, working memory and effectors form a closed reason to act to observe to revise loop; the policy perimeter gates each iteration's writes, and a budget or stop-condition bounds the cycle.
- ·Coding agents that iterate: write, run, read the error, fix (ChatGPT, Claude, GitHub Copilot).
- ·Research agents that search, read and refine a query across turns.
- ·Any open-ended task where the next step depends on the last result.
- ·Handles open-ended, unpredictable tasks a fixed pipeline cannot.
- ·Self-correcting: each observation improves the next attempt.
- ·Minimal orchestration: one core, one loop.
- ·Naturally traceable as a turn-by-turn transcript.
- ·Can loop without converging: needs hard budgets and stop conditions.
- ·Cost and latency compound with every iteration.
- ·Error accumulation: a bad early step can derail the whole run.
- ·Harder to guarantee bounded behaviour for safety-critical use.
Mapping classified by the Field Book from each agent's documented control flow.
Description, use cases and trade-offs: Field Book analysis, synthesised from established agent-architecture literature.
The corpus does not rewrite vendor documentation. Where copy is quoted verbatim it is marked as such and attributed to its source.