CORE GUIDE
Act → observe → verify loop
An agent is a control loop that chooses an action, observes the environment and verifies whether the task state actually improved.
Mental model
Every action should produce evidence for the next decision. Without observation and verification, an agent becomes a sequence of increasingly confident guesses.
Why it matters
Tool access alone does not make a reliable agent. The useful capability comes from closing the loop between intent, action and evidence so the system can detect failed assumptions, recover from errors and stop when the goal is genuinely satisfied.
01
The control-loop structure
Start from the current goal and verified state. The model selects an action; a tool or environment executes it; the runtime captures the observation; a verifier checks the expected state transition. The next decision is then continue, retry, escalate or stop.
02
Example: a coding agent edits one function
The agent edits a function, then inspects the diff and runs targeted tests. If a test fails, that observation changes the next action. If all required checks pass, the verifier can establish the postcondition instead of trusting the model's statement that the change is done.
Common failure modes
- Treating a successful tool call as proof the task succeeded.
- Skipping observation and planning the next action from stale state.
- Running without explicit termination or escalation conditions.
Engineering heuristics
- Define expected postconditions before high-impact actions.
- Prefer structured observations that can be checked independently.
- Make stop, retry and escalation explicit runtime decisions.
Takeaways
- 01Agency is a loop, not a long prompt.
- 02Actions need observations and verification to become reliable progress.
- 03Termination is part of agent architecture, not cleanup logic.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.