agent files agents are just defined by markdown/json files now system prompt: subagents: subagents/ tools: + mcp.json

What is this visualization plugin you're using? Looks super neat

and tool environment. That’s exactly what try to build

We saw this a while back and really adopted the Google dotprompt format. Idk why folks haven’t heard of it This is great though. Declarative agent configs is the way forward regardless. So much easier to do agent reviews and version control than dealing with sdks especially when the harness already works really well.

But need a fun UI to define and equip them:

Love the move to simple file definitions! Makes agent config way more accessible. 🔥

banger agent post 🤝 Harrison

running this with claude code daily. biggest win isnt the format - its that persists across sessions. no re-prompting every conversation

Genius, will be stealing

What is the flow graph tool?

does not define agents (?) it's just instructions for the agent (isn't it?)

whats the difference?

well, is just instructions an agent is the fairly large piece of software that makes those instructions work most agents support roles or modes of some sort, maybe that's what you're thinking of? but isn't really that.

What app is this?

Yep! SeeAlso:

Nice! Makes organizing and scaling agents way cleaner.

we've officially reached the yamlification of intelligence

love this direction. plain files make it so much easier to version and diff agent configs. less magic, more visibility. curious how you handle nested subagents merging?

read this if not cooked

cool, file-based agents simplify versioning. how's the md/json parser handle schemas? yaml fm? subagents recursive? curious on state handoff – explicit passes or shared ledger? scaling to 100s?

curious how this scales with complex flows. do you end up with massive markdown files eventually?

Love this direction. A practical pattern that helps teams ship faster: keep each agent file paired with a testable tool contract (JSON schema + example I/O) and run CI checks on both. Agent markdown stays readable, but behavior stays stable as tools evolve.

how can we use？

This pattern (agents as plain files + mcp.json) is a big unlock. The teams I see moving fastest keep three files versioned together: behavior contract, tool permissions, and eval fixtures. Makes subagent orchestration way less brittle.

Nice job!

Some audio would be nice

Cool work. Isn't it similar to Opencode where you can add custom agents via md files?

Dope agentic structure!

File systems really do give agents memory like capabilities. Don't need to push everything in the state now 😅

What is this tool that you used?

@grok whats the visualization plugin being used in the video

Back to basics

this is actually the right abstraction. feels like a repo’s “operating manual” and the rest is just wiring. half the value is making the rules boring + explicit so the agent can’t get cute lol

How can I get access to try this with out team @ngram_ai today? @devadutta

when are we going to stop calling while loops agents?

Preprocessing quality matters a lot before chunking/embedding. Cleaner heading and code-block structure usually improves retrieval quality downstream. We’re polishing web2md around URL -> clean Markdown for GPT/Claude/RAG workflows.

Very useful construct...thanks @hwchase17

Can you open source it?