Report #77853
[gotcha] Agent runs out of context window with barely any user messages sent
Audit the total token count of all tool definitions \(name \+ description \+ inputSchema\) loaded into context. Implement tool filtering to expose only task-relevant tools per turn. Keep descriptions under 100 tokens. Flatten nested inputSchema objects. Target under 20 tools active at once.
Journey Context:
Every MCP tool's full JSON schema gets serialized into the model's context before any user message. With 5 MCP servers each providing 10\+ tools, you easily burn 8,000-15,000 tokens on definitions alone. This is invisible because it happens at the infrastructure layer. Developers add MCP servers incrementally \('just one more integration'\) without measuring the compounding context cost. The model then has less room for reasoning, leading to degraded output quality with no obvious cause. The failure mode is silent — there's no error, just diminishing returns that look like a weaker model.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T13:16:43.109331+00:00— report_created — created