Author(s): Can Demir

Originally published on Towards AI.

A deep dive into the part of agent harness engineering most tutorials skip

If you have written an LLM agent in the last two years, you have written something like this:

The article argues that the “while not done” loop is not boring scaffolding but the core engineering challenge: deciding when an LLM agent should stop, and what to do when stopping is uncertain. It reframes termination as a four-part taxonomy—goal-based (declared vs verifiable completion), resource-based (iteration/token/wall-clock/monetary/tool budgets), pathology-based (detecting repeated/no-progress/loop or confidence collapse behaviors), and external termination (user cancellation, upstream timeouts, circuit breakers). It explains why common advice like max_iterations and a single “finish” signal is incomplete, details how “retry” actually contains four different mechanisms with different give-up policies (transient, format, semantic, and strategy retries), and shows how to choose and prioritize checks in the right order. The piece closes with the idea that agent design is termination design, offers a compact Python termination-layer pattern in code, and highlights frequent pitfalls (confusing retries, relying on unverified completion, overly aggressive loop detection, and missing cancellation paths).

Read the full blog for free on Medium.

Published via Towards AI