Agent loop

Overview

Many “agents” are just controlled loops:

• Run an iteration body
• Decide whether to stop
• Enforce budgets/timeouts
• Optionally checkpoint or interrupt

In Coevolved, this pattern is packaged as agent_loop(...), which returns a Step you can compose like any other step.

LoopPolicy and LoopState

LoopPolicy controls loop execution:

• max_iterations
• timeout_seconds
• Optional checkpoint_store + checkpoint_policy
• Optional execution_policy (budgets)

LoopState tracks iteration index, elapsed time, and the last checkpoint id.

Checkpointing and budgets in loops

The loop can checkpoint before/after each iteration and on errors/interrupts.

If a usage policy is configured, the loop checks budget before each iteration.

Stop conditions

The stop condition is a function: (state) -> bool.

Keep it:

• Explicit (no hidden globals)
• Cheap (evaluated every iteration)
• Testable (unit tests should cover stop behavior)

Next steps

ReAct agent

A concrete agent loop pattern: plan with LLM, act with tools.

prebuilt.loop reference

agent_loop, LoopPolicy, and LoopState API details.