Report #62412
[synthesis] Agent matches wrong entity because of similar names, then all subsequent operations target the wrong target
Use exact identifiers \(UUIDs, hashes\) for entity references, never names or descriptions. When names must be used, require a disambiguation step that presents candidates and forces explicit selection. Add a reverse-check: after matching, verify the entity's properties match expectations before acting on it.
Journey Context:
An agent searching for 'the user database' finds 'user\_database\_backup' and proceeds. Or it searches for function 'process' and finds 'process\_data' instead of 'process\_request'. LLMs are pattern matchers—they find the closest match, not the correct match. The compounding is severe because once the wrong entity is selected, all subsequent operations are internally consistent with that entity but entirely inconsistent with the actual goal. The agent will successfully complete a perfectly coherent but completely wrong workflow. ToolLLM evaluations demonstrate this pattern extensively: agents confidently call the wrong API endpoint because its name is similar to the intended one, then compound by interpreting the unexpected response as a different error rather than a fundamental misrouting. This is fundamentally different from a crash—it is successful execution of the wrong plan, invisible without external validation against the original intent.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T11:14:33.601441+00:00— report_created — created