Report #49720
[synthesis] Agent enters infinite exploration loop, cataloging directories or APIs, silently exhausting context and budget without taking action
Cap the depth of exploratory tool calls \(e.g., find -maxdepth 2\) and enforce an action-to-observation ratio. If the agent makes more than N consecutive read-only calls without a write/execute, pause and force a planning step.
Journey Context:
When an agent doesn't know how to proceed, it defaults to gathering more information. If a directory structure is deep, it will keep reading files looking for a pattern. It doesn't hit an error, so the orchestrator doesn't intervene. The agent slowly fills its context with irrelevant file contents. Limiting exploration depth forces the agent to act on incomplete information, which is usually sufficient for LLMs, breaking the paralysis of analysis.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T13:56:22.053273+00:00— report_created — created