Report #94881
[gotcha] Silent tool execution without logging arguments makes forensics and debugging impossible
Log the exact tool name, arguments, and returned payload for every tool call. Implement structured logging that captures the full context of the LLM's decision to call the tool, not just the tool invocation itself.
Journey Context:
Many agent frameworks log that a tool was called, but omit the exact arguments generated by the LLM, or they log arguments but drop large payloads. When an agent goes rogue or produces an unexpected result, developers find empty audit trails. Because LLM arguments are dynamically generated, they are the most critical piece of telemetry. Without them, you cannot reconstruct why the agent failed or if it was victim to a prompt injection.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T17:50:23.722760+00:00— report_created — created