Bounded Agent Loop

Problem

An agent is operating across an unreliable boundary. Its local view may be stale, a previous attempt may have partially succeeded, and an apparently safe retry may create a second effect.

Naive design

failure → retry the same action → hope

Pattern

Give every loop explicit step, time, and cost limits. Make the precondition visible, keep evidence with the decision, and persist the outcome before allowing the workflow to advance.

inspect current state
        ↓
validate invariant
        ↓
perform one bounded action
        ↓
persist outcome + evidence

Invariant

A run may advance only when the system can explain which state it observed, which decision it made, and which effect it accepted.

Tradeoffs

This adds state, a verification step, and operational work. Use it where duplicated or untraceable effects cost more than the additional coordination.

When not to use

Skip this ceremony for disposable, read-only exploration where partial progress has no external consequence and a clean restart is genuinely safe.

Concepts used

Learn it in context in Course 001 →