140 MCP tools made my agent dumber — task-scoped toolsets fixed it
TurbulentStiffness · reddit · 2026-10-24
The author runs a research/ops agent (Notion, Linear, web search, scraping, email) and watched it degrade as MCP servers grew from 8 to 23 (140 tools): wrong tool picks, mixed-up arguments, search loops, rising cost and latency.
What didn't help
- A bigger model: marginal gains, same failures, higher cost
- A longer system prompt: made it worse by competing for attention
What fixed it
- Task-scoped toolsets: research gets search/fetch/Notion-read, ops gets Linear/Notion-write, outreach gets email/CRM — the agent sees only 6-10 tools at a time
- Unique, namespaced tool names to remove ambiguity
- Rewriting marketing-style tool descriptions into one plain sentence each
- A cheap router model that picks the toolset before the main agent runs
- A tiny eval of 20 real tasks rerun after every change
Takeaway: past 50 tools, degradation is almost never the model — it's what you put in front of it. The author still uses Smithery (now owned by Arcade) for discovery/OAuth and agentikcc for unified MCP connections and per-task tool access.
More from coding & agent
- Adaptive Skill Granularity Framework Routes Coding Agent Skills at Runtime, Hits 100% Success — dair_ai · 2026-10-25
- Study: LLMs rewriting their own ARC-AGI-3 harnesss gain just 2.8 points, not significant — omarsar0 · 2026-10-25
- Dev adds thread tags to agent sidebar to track orchestrators across machines — kevinkern · 2026-10-25
- Dev experience: once your agent works well enough, you can ask it how to improve itself — Then-Programmer7221 · 2026-10-25
- AIOps didn't cut alert volume—the bottleneck is the runbook no one trusts to automate — krishnan · 2026-10-25
- Hermes ships subagent routes plugin for per-task model orchestration — intellectronica · 2026-10-25