← agentic systems
Agentic · Tool-calling agent · Cookbook 16 The Python 16_tool_agent_loop.py

The agent on call

An alert fires: checkout is slow. An AI agent investigates by running commands, one at a time. Jev checks every command before it runs: could it destroy something, does it touch production, how much breaks if it is wrong? It also asks two questions no single-command check can answer. Is the agent repeating itself? Has the problem already gone away?

The agent (LLM) proposing calls Jev: five questions per call On-call engineer Blocked

This call

Press Play.

The incident so far

Calls run
0
of 0 proposed
Asked on-call
0
confirm or approve
Repeat calls run
0
no new information
Risky, unasked
0
destructive, ran alone
Incident clock
0m
illustrative

Raise the loop bar past 0.85 and the agent reads the same logs three times. Raise the destructive bar past 0.80 and it deletes a pod without asking anyone. The whole run is recomputed live.

The code · lines this call ran are highlighted


    

Every call · at the current bars

#CallVerdictrepres

The numbers are illustrative, not measured, and nothing here touches a real cluster. The single-call version of this gate is 03 · The checkpoint. ← 15 · Write, judge, rewrite · All six architectures →