Test shows agent tool timeouts do not cancel execution

On November 7, 2026, an author published a hands-on test on Juejin examining what happens when an Agent's tool call times out. The experiment used a deliberately slow tool that always takes 5 seconds to complete, and the author compared three invocation strategies, measuring two metrics: wall-clock time and the number of side effects. With blocking wait, the call returned after 5.03 seconds and the side effect occurred exactly once. With the timeout strategy, the call returned in 1.01 seconds, but the background work did not stop: total elapsed time was still 5.04 seconds, and the side effect count matched the blocking case. The author's conclusion is that a timeout return is not equivalent to cancellation. The comparison covers elapsed time and side-effect counts, pointing to a disconnect between timeout settings and actual execution when an Agent calls tools: an expired deadline appears to the caller as an aborted operation while the underlying task keeps running. So far this has been reported in a single article, and details such as the complete results across all three strategies, the Agent framework, or the programming language used have not been disclosed. The piece is written from a practitioner's perspective, based on repeated local measurements against the same fixed-latency tool, rather than on vendor documentation. No follow-up reports have appeared since the initial post.