Skip to content

About

A small, readable kernel that teaches five agentic layers - routing, skills, orchestration, memory, and a loop. Zero API key required.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

agentic-starter

A small, readable kernel that teaches five agentic layers, plus a prompt that makes your coding agent interview you and shape the OS around your work.

No API key required for the kernel path. Nothing pretends to be a model.

First 30 minutes

  1. Install Claude Code (or use Codex / Cursor — any coding agent works, the repo does not assume a specific one).
  2. Clone this repo:
    git clone https://github.com/ishaanjayaswal-dotcom/agentic-starter.git
    cd agentic-starter
  3. Run the first command (proves the kernel with zero setup, zero API key):
    python smoke_test.py
  4. Open START-HERE.md and paste it into your coding agent to shape the kernel around your own work.

60-second path

python smoke_test.py

You should see a line like:

Your kernel booted, routed a job, and wrote a note to your vault.

That proves routing, dispatch, skills, memory, and vault write. It does not call an LLM.

Then make it yours

  1. Open START-HERE.md.
  2. Paste the block into Claude Code, Codex, or Cursor.
  3. Answer the interview. No code until you have answers.
  4. Let the agent build one real skill for your actual task.

If the interview does not happen, this repo failed you even if tests are green.

What is in here

Path Role
router.py Three tiers: deep / normal / cheap. Prints every decision.
orchestrator.py Goal -> skills -> up to 3 workers -> job record
loop.py observe / plan / execute / verify / improve
skills/ One file per skill; folder-scan registry
memory/ File store + vault writer (raw -> note)
vault/ Where notes land
tools/notes_server.py One small MCP-style tool surface
server.py GET /health, POST /dispatch, GET /jobs
smoke_test.py Zero-key green line
PATTERNS.md Doctrine: rule, failure, what to do instead
CLAUDE.md How to brief a coding agent on this repo

Optional HTTP server

python server.py

Then GET http://127.0.0.1:8765/health.

Codespaces / dev container

Open in GitHub Codespaces or any environment that reads .devcontainer/. Smoke test is the same command.

What this is not

  • Not a hosted product
  • Not a copy of anyone's private production OS
  • Not a fake brain with canned "findings"
  • Not a framework

Skills that need a model (summarise) say so and point you at START-HERE.md. Skills that do not (write_note, search_notes) simply work.

Requirements

  • Python 3.10+ recommended
  • Stdlib only for the kernel

About

A small, readable kernel that teaches five agentic layers - routing, skills, orchestration, memory, and a loop. Zero API key required.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages