Agents visual guide

The bounded agent loop

A safe agent cycle that separates planning, authorization, action, verification, and terminal states.

Editable sourceMermaidOpen full-size SVG ↗Download .mmd source

A cyclic agent flow moves from Goal to Inspect, Plan, Authorize, Act, and Verify. A successful verification reaches Complete, another safe step returns to Inspect, denied authorization reaches Stop, and failed or uncertain verification reaches Contain and Reconcile before returning to Inspect or stopping. — A dependable agent acts only after authorization and always verifies the resulting state before continuing or stopping.

Read the visual

What this shows

An agent should not jump directly from a goal to a side effect. It inspects the current state, proposes a bounded plan, checks authority, executes one authorized action, verifies the postcondition, and records a terminal outcome. Failed or uncertain outcomes move to containment and reconciliation.

Carry this forward

Key takeaways

Authorization belongs between planning and every consequential action.
Verification evaluates the real postcondition, not merely a successful tool response.
Failure and uncertainty require containment, reconciliation, and an explicit stop condition.