MCP is the wrong abstraction

Hm, not convinced the number of tools is the culprit to Tre's perf problems. The issue might be the multi-agent delegation, with different agents for "solo coding" and "document writing." Warp used to work this way, handing off between different agents to complete a task. And we realized... it performs worse. Issues with bad context handoffs, issues with one subagent trying to call tools only available to the other subagent, and noticeable latency between handoffs. We flattened everything to one agent around the time Sonnet 4 came out, and performance shot up. Still, we'll keep tool overload in mind as we build out Warp further!

Solution, create an MCP tool to get MCP tools

Very insightful. Thanks Markiplier.

MCP isn't a perfect solution for users or LLMs, but it's a damn good solution for MCP providers and MCP consumers (i.e. the IDEs). Notice how every single tool supports MCP? If you subtract MCP from the equation then each IDE, coding assistant, etc... has a separate proprietary API and it would become a extremely fractured environment. MCP has been a brilliant success for the problem it was trying to solve where it's now a universal for adding extra functionality in to any IDE, coding assistant, or anything else supporting MCP. Cloudflare's approach is also brilliant for refining it to work better inside the LLMs.

We made a protocol so the model could learn to ignore it. Based.

29:13 a benefit to the multiple model calls over writing code for all steps at once is that if a step fails, or if the choice of next step depends on the output of the previous step, the model gets a chance to pivot as needed

Generating code on demand to fulfill a users request sounds like a security nightmare we can't handle (yet)

Friend, this is your all time best video. Your ability to explain complex things in a simple but engaging way has reached a new level.

I am using MCP regulary. I think the Resources part is one of the things that is not properly understood. I regularly give specific context about Database tables by adding a MCP Resource. The MCP Resource dynamically resolves the schema of a table that I input for that resource and then adds that specific schema to the context. With that the model knows what kind of columns exist on the table with their type and exact column name. VSCode has good support for the MCP Resources in my opinion.

"You'll have an endpoint that executes specific code for each user" - What a scary thought!!!