Features / Loop detection
Stop runaway AI agent loops in seconds.
An agent that calls the same tool with the same arguments, gets the same error and tries again can run all night. ACPpotamia compares every call with the agent's recent calls and stops the session when it repeats itself too often in a short window.
How it works.
Compares what matters
The last messages, tool calls and tool results are compared, with timestamps, counters and IDs ignored, so a loop is caught even when those change.
Exact and near matches
An exact fingerprint and a similarity score catch both identical retries and near-identical ones.
Per session
Name each task with X-ControlPlane-Session-Id; one stuck task stops without stopping the rest of your agents.
Revive when fixed
A stopped session gets HTTP 429 until you revive it from the console or the API.
Questions.
What counts as a loop?
By default, five matching requests in the same session within 60 seconds. You can change the count, the window and how similar requests must be.
Does it work with the OpenAI Responses API and tool calling?
Yes. Function calls and their outputs are part of the comparison on chat completions, the Responses API and Anthropic messages.
Related features.
Kill switch
Stop one AI agent session or every agent on your team with one click or one API call. The kill switch takes effect on the next call, everywhere.
Budgets and cost control
Put a hard monthly budget on OpenAI and Anthropic spend. ACPpotamia reserves each AI agent call's cost before it is sent, so agents cannot overspend.
Audit logs
Every AI agent call logged: model, tokens, cost, outcome and timing, blocked calls included. Searchable for 365 days, with optional request and response logs.