agent.md
explorer
README.md
run.md
agent.md
model.md
tools/
web_search.md
calculator.md
agent-harness › agent.md
# agent.md — the loop
> An LLM running tools in a loop.
function run_agent(goal):
1. set history to empty
2. add system prompt
3. loop forever:
a. call model.md
b. if final → stop
c. else → run tool
Run lab
AI Cohort 2026·θby Theta

The 2-hour practice lab for AI engineers.

Hands-on labs for the patterns that make AI products actually work. Theta is the AI teacher across the curriculum. Each lab has its own AI TA — a focused expert for that domain. You write your version, the TA grades it, and a real model runs through what you built. No videos. No fluff.

New labs every month
Free during launch
Lab catalog

Browse the labs.

Hands-on labs anchored on a working principle from someone who has actually shipped agents to production. Each lab has its own AI TA.

LAB 01·agents
Shipped·~2h
THINKACTOBSERVE

Harness

withLoopyour AI TA

An LLM running tools in a loop. That's the whole field, in one sentence.

Simon Willison

The agent loop, tools, and the safety wrapping around an LLM.

You build
An agent loop that calls real tools.
Start lab
LAB 02·evaluation
Shipped·~2h
RELFMTTONEACC#1#2#3#4SCORE11 / 16

Evals

withCalyour AI TA

Evals before agents.

Andrew Ng

Rubrics, LLM judges, error analysis, and CI gates for AI output.

You build
An eval suite graded by an LLM judge.
Start lab
LAB 03·retrieval
Shipped·~2h
SYSTEM PROMPTTOOL SCHEMASMEMORYUSER GOALHISTORYTOKENS12,847

Context

withVerbayour AI TA

It's context engineering, not prompt engineering.

Andrej Karpathy

The window is the program. RAG, embeddings, chunking, hybrid search.

You build
A live RAG pipeline over a real doc set.
Start lab
LAB 04·protocols
Shipped·~2h
fsgitslackdbnotesHOSTmcpPROTOCOLjson-rpc 2.0

MCP

withWireyour AI TA

MCP is USB-C for AI agents.

Anthropic

JSON-RPC, tools, resources, prompts — and the seven sins that break production.

You build
A full JSON-RPC MCP session.
Start lab
LAB 05·training
Shipped·~2h
FROZEN · WΔWrank · rW' = W + B·A

Fine-tune

withForgeyour AI TA

Most fine-tuning problems are prompting problems in disguise.

Forge

When to fine-tune, when not to — and how to actually train a small open-weight model with LoRA.

You build
A LoRA adapter on a 1.5B model, measurably better than base.
Start lab
LAB 06·agentic IDE
Shipped·~2h
PLANEDIT/plan/clear/agentsplan · execute · verify

Claude Code

withPairyour AI TA

Senior engineers don't type first. Neither should the model.

Pair

Plan mode, slash commands, CLAUDE.md, subagents, hooks, skills. The harness Anthropic shipped for engineers, taught at engineering speed.

You build
A real PR shipped end-to-end with plan mode, subagents, and a code-review pass.
Start lab
LAB 07·real-time AI
Shipped·~2h
YOUVADSTTLLMTTS200ms800msAGENTlisten · think · speak · before the wall

Voice

withEchoyour AI TA

Voice is a real-time relay race, not a request-response.

Echo

VAD, STT, the realtime model, barge-in, telephony, eval. The whole stack for shipping a voice agent that doesn't sound like 2009.

You build
A working voice agent that hits sub-500ms with barge-in over a real phone number.
Start lab
LAB 08·agentic interfaces
Shipped·~2h
INTENT{type:Cardtitle:"Sales"data:[3,7,2,9]cta:"Open"}SalesOpenintent → typed components → live UI

Generative UI

withCanvasyour AI TA

LLMs don't return UIs. They return intents. The runtime returns the UI.

Canvas

Component contracts, streaming hydration, WebMCP (W3C Draft), Google A2UI (shipped Dec 2025), MCP Apps. Where chatbots become apps.

You build
A generative UI agent that streams typed components — not strings — into a live interface.
Start lab

Frequently asked questions

Theta is your AI teacher across the whole curriculum. In machine learning, θ is the symbol for the parameters being learned during training — so we named the teacher after the thing being shaped. Each lab also has its own AI TA (Loop for agents, Cal for evaluation, Verba for context, Wire for protocols). They're focused on their domain; Theta keeps the through-line.

Two hours, designed for one focused sitting. The progress saves to your browser, so you can split it across two sessions if you have to — but the lab is shaped for one go.

Yes — these labs assume you're an engineer. They don't teach Python or Git. They teach the specific patterns of building AI products: agent loops, evals, RAG, MCP. If you've shipped any production code, you're qualified.

Two things. First, format: every lab is two hours, finishable in one sitting, with a clear artifact at the end. Second, AI-graded worksheets: before you read the canonical answer, you write your own, and your lab's AI TA reads what you wrote and gives specific feedback. No videos. No multiple choice.

No. The artifact is the certificate — you finish each lab having built a working thing you understood. If your employer wants proof, share the lab URL plus what you built.

Free during the launch period. Future labs may be paid, but everything currently shipped stays free.

New labs ship roughly once a month. The next ones in the queue are Tools, Production, Safety, and Multi-agent. Theta stays the teacher; each new lab gets its own AI TA.

An independent engineering studio — Theta Studio. The curriculum is anchored on canonical sources (Karpathy, Anthropic, Ng, Husain, Willison) and built with the discipline you'd want in production systems, not in vibe-y demos.