agentsop-bounded-loop

To stop infinite agent loops, you need an explicit termination contract using a counter budget, checkable exit predicates, and stagnation detection outside the language model's control. This prevents runaway costs, rate-limits, and unusable failures by fast-failing repeated errors and routing to an escalation path with preserved state.

Multi-agent delegation ping-pong happens when handoff loops lack explicit termination mechanisms. Implementing a framework-agnostic termination contract with shared persistent state updates and pure-Python predicates detects stagnation, enforces iteration budgets, and breaks unbounded retry cycles.

The best way to bound cyclic agent pipelines is enforcing a progress witness requiring each iteration to change state measurably, combined with an iteration counter and an escalation branch. This framework-agnostic approach fast-fails when outputs repeat instead of endlessly retrying.

Raising recursion_limit, max_iter, or max_turns to fix agent retry errors is an anti-pattern that increases costs and delays failure. Instead, refactor your workflow to add an iteration counter, progress witness, stagnation exit, and an escalation branch that preserves state.

Yes, pure-Python predicates can terminate self-healing code loops by evaluating shared persistent state updates outside the language model's control. This framework-agnostic termination contract ensures loops exit gracefully via a human or outer fallback when stagnation is detected.