Agent-Reflex: Teaching Language Agents to Act Reflectively

The short version

The problem

Tool-using assistants can get stuck when a service times out, returns misleading information, or a user corrects their request. Most are trained on smoother interactions, so they may repeat a bad action or miss the correction.

The idea

Agent-Reflex teaches recovery from experience: it compares successful and failed attempts in the same tool environment and distills reusable advice, such as checking an item’s identity before updating it. The agent practices fixing a mistake in place, restarting a flawed plan, and responding to a user’s correction; later, training uses that advice to guide which actions the agent should learn from.

What it got

The trained 8-billion-parameter agent reaches an overall score of 58.6 across five long tool-use tests and 29.1 on a test with noisy tools and user interactions. It beats the compared agents of similar size on these overall results.