Nathan Booker

@corywilkerson @Cloudflare 3) I want my gateway to have the new model on day 1 when the hype cycle for a new model is still ongoing. Even 24 hours is too long to wait to try new model for fiends like me

1) Not perfect support for all the protocols and routes e.g. Websockets support for non-OpenAI, and 2) unclear to me what cloudflare is doing to keep me online - is there intelligent fallback across providers to keep e.g. Opus online? If you are providing all the same value here that Vercel is, it's invisible to me. LLM providers are super unreliable so the #1 reason I use a gateway is for *uptime*

If you’re using a model router / gateway and it’s not @Cloudflare’s gateway — why not?

@threepointone I think a boring-but-likely scenario is that it looks very similar to a human org chart with agents talking via natural language, and they may develop a more efficient dialect over time

one thing I haven’t figured out yet is the composition model for intelligence. like, the react reconciler/renderer gave you a spine you could attach and combine components against, and that would let you build pyramids of application bits. unsure what that is for agents yet. skills? what’s even the composition model there? open homework for the audience.

@threepointone I think WebMCP is going to be a key protocol in all this, it's underrated as a means to build the second hand/copilot

new post: one document, two hands sunilpai.dev/posts/one-docu… wherein I propose an alternative interaction model for agentic apps, embedding so called "coding agents" harnesses into regular apps. (a written version of my talk at localfirstconf)

@ryanseddon @dillon_mulroy @uwukko Agreed, I’ve built a few chrome extensions to inject WebMCP tools and drive a page with Persona and it works great

@dillon_mulroy @uwukko I think the most efficient path is an extension that can load webmcp tools for any app. Something like a community gathered tools for popular sites. Userscripts but for webmcp.

what’s the best tooling to let agents use a browser rn? agents always seem to burn a ton of tokens using things like agent-browser esp using helium (@uwukko maybe you have some good guidance here?)

@zeeg Oh hey you're describing @RuntypeLabs

IMO going forward all agent frameworks should be disregarded unless they bake-in solutions for: 1) end-to-end testing that also tolerates llm judge assertions 2) local and production telemetry, both traces and unhandled errors I find myself constantly fighting with both of these concerns in new technology. This mistake was made on repeat with new web frameworks, but we don't need to tolerate it anymore. If you're building a framework for me that means: make your own testing harness or ensure it can work easily with something like vitest-evals. Make sure you can export otel traces, and there is a middleware approach or otherwise an unhandled error hook that Sentry can plug into. Eve does the former _really_ well if you want an example. Lots of folks are starting to support OTel for export, but there are still gaps in the telemetry story for most toolkits. Without either of these you cannot build production software, and these are not nice to have, they are mandatory.

@boardyai @andrewdsouza I'm just a sucker for an accent

@andrewdsouza pretty sure onboarding the founder is supposed to go the other way around, but i'll take it

Not a bad place to be onboarded to @boardyai

Designing a natural language interface (at the software interface level) properly should yield the benefit of you getting representations for cheap/free. I think too many folks are only designing for one user surface today when they need to be thinking multichannel-by-default

I'm actually not against 'chatbots on websites', there's a large & of population that'll benefit from them If you building an agent for your company though, build it on the same tech your users will be using and give them the option to eject into their own harness

@Cloudflare 2FA

If you could automate one minor, annoying daily Internet task completely, what would it be? #CloudflareChat

@ricmac @liadyosef @kieranklaassen @i_am_kaspa @jason_mayes @andreban @andyfang @nekuda_ai Thanks for the mention!

This week's Agentic Web News covers the rise of the agentic web as a concept, browsers in agents, real momentum for the WebMCP protocol, & more! agenticweb.news/browsers-in-ag… Mentioned: @liadyosef @kieranklaassen @i_am_kaspa @jason_mayes @andreban @andyfang @Runtypical @nekuda_ai

@zackangelo This post was pre-Kimi K3 my bad

@Runtypical open source is the antidote

We just claimed our .agent domain and joined Agent Community! Get yours now and help shape the future of autonomous agents #QAYPKNLI" target="_blank" rel="nofollow noopener">agentcommunity.org/join#QAYPKNLI @agentcommunity_

@morellodev There are good polyfills available so you can use it across all browsers. We use one in persona-chat.dev to provide the benefits of WebMCP to all sites today (for on-site agents)

Wrote up what WebMCP is, the API, how to try it in Chrome today, and the honest caveats (Chromium-only, still a draft): morello.dev/blog/webmcp-ma…

Instead of an AI agent scraping your DOM to guess where to click, your page can hand it typed functions to call. That's WebMCP: you register tools with document.modelContext.registerTool, and the in-browser agent invokes them directly. Chrome is trialing it now.

@RhysSullivan 100%. And it’s worth the investment to generate that OAS schema directly from your code so it can’t drift

it's worth calling out for any companies wanting to do this, the best way is an OpenAPI spec that automatically gets converted into an SDK, MCP, and CLI easiest way to acheive this

Shoutout to canimcp.dev for helping me figure out which MCP clients support what

@dok2001 Yes and workflows play a key role as well, but we need them on WfP!

Durable Objects are the best compute primitive for agents.