How many tools can an agent reliably select from?

Research shows accuracy degrades significantly past 20-30 tools and drops below 50% at 50+ tools. Vercel found removing 80% of tools raised success rate from 80% to 100%.

What is the 12-factor-agents framework?

An open-source guide (23k GitHub stars) listing engineering principles for reliable agent software, including keeping tool counts small, managing context explicitly, and building stateless agents.

← Back to dashboard

clawsmith.com/signal/agent-over-tooling-kills-accuracy-50-plus-tools

⚠ IssueWide Openai_agent_mcpLive

Loading agents with 50 or more tools drops tool selection accuracy below 50 percent

Research and production post-mortems show agent accuracy degrades sharply as tool count increases. At 50+ tools the agent picks the wrong tool more than half the time. Vercel removed 80% of their text-to-SQL agent tools and got 3.5x faster with 100% success vs 80% before. The 12-factor-agents GitHub repo (23k stars) lists 'do not overload with tools' as a core reliability principle. Developers pile on more tools when agents fail, making the failure worse. No standard tool relevance scoring layer exists.

Product Idea from this Signal

A MCP server that dynamically routes tool selection for AI agents at runtime

25.5k ▲

AI agents loaded with 50 or more tools suffer severe accuracy degradation because the model must process every tool schema upfront, overwhelming the context window and confusing selection. There is no production runtime layer that intelligently surfaces only the tools relevant to the current task turn, measures selection accuracy, and hot-swaps the active toolset without restarting the agent. This MCP server sits between the agent and its tool registry, scoring relevance per-turn and injecting only the top-N tool schemas, keeping accuracy high as the registry grows.

ai-agentsmcptool-routingcontext-optimizationdeveloper-tools

Competitive29 leadsView Opportunity →