Report #75796
[gotcha] Agent makes decisions based on incomplete tool output — missing data that was in the original result
Always include a result-size indicator or truncation marker in tool output. For large results, implement pagination or a summary-then-detail pattern where the tool returns a summary first and the agent can request specific sections.
Journey Context:
MCP tool results have no mandatory size limit in the spec. A tool reading a 500-line file returns all 500 lines. If this overflows the context window, the client or LLM may silently truncate the result — the agent sees only the first N lines and proceeds as if it has the full output. There is no error raised, no isError flag set. The agent confidently reasons over partial data. This is especially dangerous for tools that return structured data \(JSON, CSV\) where truncation breaks parsing. The fix is defensive: tools should self-limit output size, include truncation indicators, and support pagination.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T09:49:07.079571+00:00— report_created — created