New framework helps developers test and fix LLM agents when tools fail

Developers working with large language model agents that rely on external tools now have a better way to catch and resolve problems before they reach users. The new AgentCheck workbench addresses a critical gap: most testing assumes tools always work perfectly, but in real deployments, services time out, return outdated data, or get misconfigured.

The Problem With Tool-Using Agents#

AI agents powered by LLMs often depend on calling external tools—APIs, databases, services—to complete tasks. Current evaluation methods test agents under ideal conditions, where every tool response arrives correctly and on time. But in production, things break. A service might time out. Data might be stale. A tool's description might be corrupted or poisoned. When this happens, developers have no reliable way to test whether their fixes actually work before deploying them.

How AgentCheck Works#

AgentCheck turns this into a controlled testing loop. The system runs an agent against real tools while recording every response. Then it automatically replays the agent's execution with intentional faults injected—the framework includes 12 different failure types to simulate common real-world problems.

The replay is clever: matching tool calls come straight from cache to ensure reproducibility, while new calls go live after the agent diverges from the original path. This allows developers to toggle a potential fix and re-run the exact same failure scenario to see if it resolves the issue.

Scoring happens in two stages: hardcoded pass/fail rules handle clear outcomes, and an LLM judge evaluates more interpretive results, validated against human feedback.

What The Tests Revealed#

Across five different agents tested, performance varied widely. The strongest agent handled 105 out of 120 test scenarios successfully, while the weakest managed only 77. Most failures weren't dramatic crashes but rather silent errors—the agent confidently used incorrect tool outputs without raising an alarm.

On that weakest agent, a simple retry mitigation improved timeout handling dramatically, raising success from just 30% to 100%. However, the same fix barely helped with stale-data problems, which stayed stuck near 30-40% success regardless.

Why This Matters#

AgentCheck makes failure modes repeatable, measurable, and verifiable before anyone relies on the agent in production. For developers, that means catching problems earlier and validating fixes with confidence rather than guesswork.