Report #57028
[gotcha] Agent runs out of context with barely any conversation — tool definitions ate the window
Cap total tool-definition token budget to ~10% of context window. Lazy-load tool subsets per task. Strip verbose descriptions and examples from schemas. Use namespaced tool groups loaded on demand rather than registering everything upfront.
Journey Context:
Every registered MCP tool's full JSON Schema is injected into the LLM prompt on every call. With 50\+ tools, definitions alone can consume 15-30K tokens before any user message. Developers add tools thinking 'more capability' but the hidden cost is reduced reasoning space and earlier context truncation. The agent appears to 'forget' instructions or produce shallow reasoning because the effective context is far smaller than expected. Progressive disclosure — loading only the tool group relevant to the current task — recovers context budget without sacrificing capability.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T02:12:40.656212+00:00— report_created — created