Chirag Jain

Commentary on the possibilities and progress of AI to get things done. I design, build and deploy agentic workflows for knowledge work tasks.

[Chirag Jain @chirag2653]

@lydiahallie Better harness tools to spawn and manage headless claude -p sessions similar to subagent tool. I understand each session is a new OS process but it can unlock lot of use cases.

[Chirag Jain @chirag2653]

Managed Agents feature request — the hosting infra is exactly what I needed. Sessions, containers, async execution, file persistence. But here's the gap: I've built a production agent locally in Claude Code using skills with context: fork, allowed-tools, hooks, subagents with frontmatter, CLAUDE.md layering. None of that orchestration transfers to Managed Agents. The Managed Agents harness only honors name and description from skill frontmatter — Claude Code has 13+ fields. No hooks, no forking, no subagent .md files.

The engineering blog says Managed Agents is a "meta-harness" that can accommodate different harnesses including Claude Code. So: can we get a harness: "claude-code" option when creating a session? Container boots, mounts my .claude/ directory, Claude Code CLI runs headlessly inside it. Full skill frontmatter, hooks, subagents — everything works.

Right now the choice is: rich orchestration locally OR managed cloud infra — but not both.

@trq212 thoughts?

[Chirag Jain @chirag2653]

@trq212 Do Claude Managed Agents support Hooks and Subagents? Forked Skills?

[Chirag Jain @chirag2653]

Every knowledge worker will have this. The "PA" is the new "PC." Just like everyone got a personal computer in the 90s, everyone gets a personal assistant from this year onwards 🤞

[Chirag Jain @chirag2653]

@noahzweben The biggest issue I face with using Claude Code in cloud is loss of my environment setup on my machine particularly all the CLI tools I have installed. Can you let user's easily mirror their local Claude Code CLI tools setup in a custom environment configuration in the cloud?

[Chirag Jain @chirag2653]

Interesting move from @OpenAI

Claude Agent SDK = "Here's a powerful agentic loop library. Host it yourself, sandbox it yourself, expose it yourself."

OpenAI Responses API = "Here's a powerful agentic loop API endpoint. We host it, sandbox it, and expose it. Just call us."

OpenAI is making the infrastructure disappear. You don't need Railway, you don't need e2b, you don't need Express. The API IS the deployment. The container IS the sandbox. The skills ARE the deployment artifacts.

The trade-off is lock-in and control. With Claude's approach, you own everything. With OpenAI's approach, you're renting their platform — but shipping 10x faster.

[OpenAI Developers @OpenAIDevs]

We're introducing a new set of primitives in the Responses API for long-running agentic work on computers.

Server-side compaction
• Enable multi-hour agent runs without hitting context limits.

Containers with networking
• Give OpenAI-hosted containers controlled internet access to install libraries and run scripts.

Skills in the API
• Native support for the Agent Skills standard and our first pre-built spreadsheets skill.

https://t.co/vK9fbhHQdq

[Chirag Jain @chirag2653]

@mckaywrigley I really believe you are among the few people who take a lot of effort in communicating the value of having something as powerful as Claude Agent SDK (now with Agent Swarms) to build agents. Would love more content from you and demos on this topic.

[Chirag Jain @chirag2653]

@matt_elevenlabs 11B

[Chirag Jain @chirag2653]

@JackWoth98 @geminicli Can we have something to write a Skill Suite for a big e2e workflow with an orchestrator skill to stitch them in a sequence?

[Chirag Jain @chirag2653]

💡 Pro tip for AI agent power users:

Store your API keys as OS-level environment variables, then write a shared skill script that reads from them.

Result? One setup, works across:
→ Claude Code
→ Gemini CLI
→ Cursor
→ Codex
→ Any future agent

No re-entering keys per tool. No hardcoding secrets in configs.

Your skills become portable infrastructure — the agent is just the runtime.

[Chirag Jain @chirag2653]

@felixrieseberg I thought slash commands and skills were merged? Clarification please

[Chirag Jain @chirag2653]

Hey @rauchg - every Next.js page already exports metadata: title, description, canonical URL, OG tags.

That's literally YAML frontmatter waiting to happen.

Right now when AI agents need website content, they all do the same wasteful dance:

🔄 Hit the URL → Firecrawl strips HTML → returns markdown → agent reads it

Every agent. Every request. Same conversion. Over and over.

What if next build had an optional flag that also outputs a /skills folder? Every page as markdown, existing metadata as frontmatter, auto-generated https://t.co/y1p6PdxkQ3 as index. Served from the same Vercel CDN, always in sync with the site.

No new content to maintain — it's a second render target from the same source. HTML for humans, markdown for agents. Same deploy.

Next.js is already THE framework AI agents build with. Why not make it the first framework AI agents can natively read too? 🚀

[Chirag Jain @chirag2653]

🧵 We're entering the age of "skill folderification" of documentation.

The primary reader of your docs is no longer a human browsing a website. It's an AI agent grepping through files.

Notion, Confluence, - all optimized for human browsing. Rich formatting, sidebars, search bars. But when an AI agent needs
your docs, all that gets stripped to markdown anyway.

You're paying for a rendering layer the primary consumer doesn't use.

What agents actually need:

📄 A https://t.co/y1p6PdxkQ3 index file - scan titles to find what's relevant
📋 YAML frontmatter on every page - decide relevance without reading the body
📝 Markdown body - load only when needed

Progressive disclosure. Not for humans clicking through pages - for agents managing context windows.

This isn't theoretical. The "agent skill" standard already works across multiple AI agents - Claude Code, Cursor, Codex, Gemini CLI, and
more.

One folder. Markdown + frontmatter. npx skills add. Symlinked everywhere.

Your docs become a dependency, not a website. 📦

The filesystem is the new docs platform.

[Chirag Jain @chirag2653]

@zarazhangrui "Coding is just a means to an end". Well said. And the end is usually a solution to a problem.

[Chirag Jain @chirag2653]

We need a universal package manager for MCP servers like what we have now for Skills with @vercel npx skills.

[Chirag Jain @chirag2653]

I love your video demos; they look really cool and informational. Thank you for it.

I have this question; maybe you can help me. What is the video making workflow? Is it - Recording the audio first and then adding the video clips as B-rolls or
First recording everything as a Loom-style video and then editing.

[Chirag Jain @chirag2653]

the @vercel https://t.co/5UL6OOeY7x CLI is genius because it can own discovery. If it indexes skills across GitHub/GitLab (and eventually the open web via https://t.co/y1p6PdwN0v + metadata), “how to do X” stops being blog search and becomes “run find; install; agent does it.” That creates an SEO-like game: devs, companies, and marketers will optimize to rank in npx skills find—and whoever controls that index controls the workflow. @rauchg

[Chirag Jain @chirag2653]

What a SKILL.deps.yaml Could Look Like

requires:
runtime:
- python: ">=3.8"
cli:
- python3: "*"
api_keys:
- OPENAI_API_KEY: "optional" # enables full mode
- XAI_API_KEY: "optional" # enables X research
agent_capabilities:
- WebSearch: "required" # Claude Code built-in
filesystem:
- ~/.config/last30days/: "write"

[Chirag Jain @chirag2653]

Skills are instruction manuals for agents.

But manuals assume you have the tools to execute the workflow as precibed by the skill author.

What if every SKILL folder had a DEPENDENCIES.json file?

{
"cli": ["ffmpeg", "yt-dlp"],
"mcp": ["@anthropic/mcp-filesystem"],
"scripts": ["./scripts/convert.sh"]
}

Agent reads deps → installs missing → skill works reliably.

npm has package.json
Python has requirements.txt
Skills have... nothing?

@bcherny thoughts?

[Chirag Jain @chirag2653]

@rauchg https://t.co/5UL6OOfvX5 CLI tool is becoming the de facto way to install agent skills - centralized ~/.agents/ storage + symlinks to any agent (Claude Code, Cursor, etc.)

We need the same for MCP servers.

Imagine: npx mcp add <github-url> → scans for mcp config → installs to all agents

Ship it? 🚀