Outcome Monitors project cover

Projects

Outcome Monitors

Recovery affordances that let an agent notice and repair tool calls that failed silently.

#AIAgents#TrustworthyAI

Overview

A tool call that returns a well-formed response the agent cannot act on is worse than one that raises: the agent proceeds as if the step succeeded, and the error surfaces several steps later, if at all.

Approach

Outcome monitors sit after the call and check the result against what the step was supposed to achieve, giving the agent a place to notice the failure and a concrete affordance for repairing it.

Papers

Outcome Monitors: Recovery Affordances for Silent Tool Failures S. Panthi, R. Abdelfattah · arXiv preprint · 2026 [LINK]

Write-ups

Recovery Tools Help Agents Respond to Corrupted Tool Results Agents trust plausible but wrong tool output; a monitor helps only when its receipt names something the agent can do next.