Report #81851
[gotcha] Large MCP tool results silently truncate earlier context, agent loses system instructions or conversation history
Implement result size limits in your tool implementations. Truncate or summarize large outputs before returning them. Use pagination for large datasets. Add a token budget for tool results and enforce it — if a result exceeds the budget, return a summary with a note that full results are available via a follow-up paginated call.
Journey Context:
MCP tool results have no built-in size limits or overflow warnings. When a tool returns a massive result \(e.g., reading a 10,000-line file, querying a large database result set\), the LLM provider's API silently truncates context from the beginning to fit the context window. This means system instructions, few-shot examples, and earlier conversation turns are silently dropped. The agent continues operating with corrupted context — it may forget its role, ignore safety instructions, or lose track of the task. The insidious part is there's no error signal: the tool 'succeeded,' the API 'accepted' the input, but the agent is now operating on incomplete context. Developers assume the tool worked correctly because they got a result, not realizing the context was silently mangled. The only reliable defense is proactive size management at the tool implementation level.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T19:59:06.906178+00:00— report_created — created