I Gave One MCP Agent 2,000 Tools (Here’s What Broke)
🔗 Link do vídeo: https://www.youtube.com/watch?v=jyYDHJAcfak
🆔 ID do vídeo: jyYDHJAcfak
📅 Publicado em: 2025-12-10T21:13:55Z
📺 Canal: Arseny Shatokhin
⏱️ Duração (ISO): PT19M56S
⏱️ Duração formatada: 00:19:56
📊 Estatísticas:
– Views: 429
– Likes: 17
– Comentários: 3
🏷️ Tags:
🐙 Our framework starter template: https://github.com/agency-ai-solutions/agency-starter-template/tree/main
🐝 Link to our framework: https://github.com/VRSEN/agency-swarm
👨💻 If you're a developer / agency and want to build your own AI solutions, join our FREE Skool community. You'll find support on every step of the way: https://www.skool.com/agency-ai
🚀 Try out our platform when you’re ready to deploy your agents to production: https://agencii.ai
Timestamps:
00:00 – Intro
01:25 – How It Works
02:29 – Building the Agent
10:11 – Testing the Agent
15:48 – Creating Reusable Tools (76K to 8K tokens)
18:38 – Production Recommendations
About:
In this video I push AI agents to the limit by giving a single executive assistant agent access to *2,000+ MCP tools* — and letting it *discover, connect, and compose them programmatically* in real time. Instead of hard‑coding tools up front, we use *programmatic tool calling + progressive tool discovery* so the agent can generate its own code, search a massive MCP server library, and build custom workflows on the fly.
You’ll see how I:
– Start from an *AI agent template* in Cursor and wire it for *MCP code execution*
– Use a scraped library of *2,000+ MCP servers* (Notion, Gmail, Slack, etc.) as a searchable tool index
– Let the agent *self‑connect* to Notion, Gmail, and Slack MCP servers with authentication
– Have the agent *read real project data* and draft a client update email across multiple tools
– Turn these ad‑hoc workflows into *reusable “skills”* (code files) to cut token usage by ~10x
– Combine *programmatic tool calling for development* with *direct tool calling for production reliability*
If you’re building *production AI agents, AI assistants, or autonomous workflows* and want to understand where *MCP, tool discovery, and code‑based tool execution* actually make sense (and where they break), this video walks through the full process, trade‑offs, and a pattern you can start using today.