This is a solution seeking a problem. Why would we want a single reliable agent to have access to 1000s of tools? When would this ACTUALLY be valuable or useful.
What IS useful and offers value is having an agent which accesses 1 or 2 tools but always uses those tools accurately and correctly 99.9+% of the time.
I've seen tons of MCP companies who are offering 1000s of wrapped HTTP APIs as MCP tools, which is note very easy to implement, but in reality, it's totally fucking useless for enterprise use-cases which need to work reliably, in a secure, repeatable fashion.
Any chump can rig an MCP client to 20 tools, but then watch your agent fail again and again and again.
Basically, this is a bad idea for a business and I'd personally suggest pivoting to something that focuses on ensuring a single agent works reliably, provides guardrails, evaluation, security etc. This is the real challenge to solve.
I'd say the value of Strata is more than just one MCP Server, it's a strategy that can apply on AI Agent interacting with tools or even other AI agents.
in your example, one agent access 1~2 tools, what if an enterprise have 1000 use cases and need 1000 AI agents as you said. How you let your system intelligently dispatch the correct AI agent to the correct use case? (or will OpenAI create 1000 customGPT and let user to choose which one? ） instead you need a way to intelligently dispatch user query to different sub-tools or sub-agent, depends on how you design. But the essence of Strata I believe is rely on model's reasoning to minimize context window + maximizer relevant info. and then solve the user query.
Security is another level of topic and I think it's not related to this approach.
The reality is that enterprises don't need 1000s of tools. I've worked for enterprises, and I've also been involved heavily with horizontal and vertical agent platforms.
The idea of 20000 MCP tools sounds good on paper, but the reality is that the enterprise wants an agent that just connects to Jira and ServiceNow and actually works. They don't care about 1000 randomly shitty MCP servers built by random people online which they can't trust.
What IS useful and offers value is having an agent which accesses 1 or 2 tools but always uses those tools accurately and correctly 99.9+% of the time.
I've seen tons of MCP companies who are offering 1000s of wrapped HTTP APIs as MCP tools, which is note very easy to implement, but in reality, it's totally fucking useless for enterprise use-cases which need to work reliably, in a secure, repeatable fashion.
Any chump can rig an MCP client to 20 tools, but then watch your agent fail again and again and again.
Basically, this is a bad idea for a business and I'd personally suggest pivoting to something that focuses on ensuring a single agent works reliably, provides guardrails, evaluation, security etc. This is the real challenge to solve.