I feel I don't do real coding 😅
I had a great time at @freestyle_dev's Bugs & Beer night.
You can RSVP to the next one here: https://t.co/3IxLT2FsXB
Pani Puri Takes #1:
We sometimes spend weeks polishing landing pages.
That’s like obsessing over the fancy lights for your pani puri stall before figuring out the pani recipe.
First make your first customer say:
“Bhaiya, ek aur dena.”("Bro, one more please")
Then worry about the fancy lights.
Me and @OneShotViber were one of the winners @nozomioai hackathon held last weekend.
We won the @insforge track where our agent hosted our frontend, DB, functions and backend on insforge using their agent first cli and MCP server.
Our Idea was an app that lets companies create custom AI playgrounds.
@benswerd Building custom v0 style playgrounds for dev tools to replace docs agent.
Users can just spinup custom starter apps instead of copying skills, MCP , repos or docs.
Dev tools get a rich flow of feedback on what their users want.
Made this for copilotkit.
https://t.co/obCYSLijLI
AI playground to spin up custom starter apps for @CopilotKit and @AGUI_Protocol
It's lovable style app where a coding agent made with @mastra , uses @daytonaio sandboxes to help you create custom generative UI apps using AGUI.
Now you don't have to clone repos, install skills, wire MCPs just to get a POC for your idea.
Reply for access to the app.
AI playground to spin up custom starter apps for @CopilotKit and @AGUI_Protocol
It's lovable style app where a coding agent made with @mastra , uses @daytonaio sandboxes to help you create custom generative UI apps using AGUI.
Now you don't have to clone repos, install skills, wire MCPs just to get a POC for your idea.
Reply for access to the app.
"MCP is dead" 💀
Meanwhile: 32M downloads *every week*
Build MCP Apps and deploy them inside your app with CopilotKit: https://t.co/bPsDtEbS7p
Open source - built with @Mastra, @e2b, @manufact, and CopilotKit.
The urge to satisfy my own hubris, that the first idea or architecture i came up with is the goat and be attached to it like its something precious.
Biggest time waster, i guess being close to a user and having hard deadlines help.
AG-UI protocol has become as an industry standard📈
The Agent-User Interaction protocol allows you to connect AI agents and frontends.
AND it just crossed 2.5M weekly downloads!
With adoption/integration by Google, AWS, Microsoft, Oracle, LangChain, Mastra, Pydantic AI, CrewAI & so many more.
As well as SDKs in .NET, Golang, Kotlin, Java, Rust and more all contributed by the community.
How it works & more 👇
Introducing: Lovable for MCP Apps 💗
Build full MCP Apps from a single prompt, in real time.
A CopilotKit chat drives a @Mastra agent that can spin up @E2B sandboxes running the @Manufact mcp-use library.
It's open source 👉 https://t.co/bPsDtEbS7p
Just saw @mariusbl Marius Buleandra (Anthropic) talk on “Demystifying Evals for AI Agents”, hit every pain point I’ve been feeling
Key Takeaways:
> Outcome-first eval >instead of rigid transcript based (let agents solve creatively!)
> Capability evals + regression evals, to test new and past features of the agent
> Re-test the full stack as models change: agent harness + prompts + model. With new models the harness could be the limiter
> Public evals are a solid base — adapt and add custom ones to march your agent architecture + env
> Eval environments are critical: design for clean resets every trial or you’ll get noisy correlated failures (env + harness must play nice)
> Mix deterministic code checks + LLM Agent judges + Human evals
> The LLM Agent judges should be built with clear structured rubrics; calibrate them by testing against human graders on the same tasks
> Custom tooling to speed up human oversight of the whole process. Ex. to see results, make edits etc
Anthropic blog link in next tweet