Language model agents struggle to spot plausible tool errors
Loud Failures, Quiet Failures: Fault Detection and Recovery in Tool-Using Language Model Agents
Artificial Intelligence
Summary
Language model agents use external tools to solve multi-step tasks but often miss when these tools give believable but wrong answers. The authors tested several language models by introducing different types of tool errors and found that agents noticed obvious errors almost all the time, but missed many silent failures that looked correct on the surface. Models designed to reason more deeply did not recover from errors any better. The study also showed that agents sometimes repeat the same tool calls unnecessarily and do not improve detection by simply asking them to check results.
What this means in practice
- •For ai system engineers: Improve multi-tool AI agents by focusing on detecting silent tool failures that look correct rather than only explicit errors.
- •For automation platform developers: Design more robust multi-step automation workflows that handle subtle tool output errors to reduce task failures in deployed AI services.
Authors
Obada Kraishan
Abstract
Tool-using agents are usually scored on whether they finish a task while the tools work. Deployments are less forgiving: services time out, endpoints disappear, parameter names change, and results come back well formed but wrong. Prior work has shown that language models over-trust tool outputs that fail silently; we ask how that over-trust plays out across the stages of failure handling in multi-turn agents. Wrapping the executable environments of an established function-calling benchmark in a fault-injection layer, we inject one of four typed faults at a controlled point in the trajectory and record whether the agent notices, changes plan, recovers the task, or repeats itself. Six models from three families, half of them reasoning variants, ran 1,920 trials over 24 multi-step tasks. Agents treat a failure as a problem in 91.3% of trials when the tool returns an explicit error, but in 58.8% of trials when it returns a plausible wrong value, against a 26.8% rate of reporting problems when nothing was wrong. Reasoning models are not better placed: paired against instruct siblings, they notice less (-9.3 points, p < .001) and change plan more (+10.4 points, p < .001), and recovery is unchanged (p = .512). Because agents are stochastic, two fault-free runs of the same task end in the same state only 63.3% of the time; against that baseline, only a missing tool clearly lowers recovery (39.9%), while timeouts, schema drift, and corruption stay within run-to-run variation. After a fault, agents return to the same tool three or more times in a row in up to 22.2% of trials, though strictly identical repeats are rare. A prompt line asking the agent to check each result did not move detection. Agents respond to the error channel rather than to the content of what a tool returns, so failures that stay inside the expected format pass through.