The promise
You build a real AI agent from scratch — on your own machine, with a local model and no API key. Not theory. Not a wrapper around a framework. You write the loop, the tools, the memory, and the trace yourself. Then you give it real capabilities, ship it behind an API, connect it to MCP servers, teach it reusable skills, and engineer reliable autonomous loops.
Six workshops. One coherent journey. Every session ends with something you built.
"If you can't build it from scratch, you don't understand it."
The arc
Build → Empower → Ship → Connect → Extend → Engineer
CORE (launch now)
W1 Build It streaming agent + 3-file memory
W2 Empower It filesystem/terminal tools + safety
W3 Ship It web/email tools + FastAPI + client
ADVANCED (after cohort 1)
W4 MCP agent as a client to any server
W5 Skills package your own capabilities
W6 Loop + Prod loop engineering, eval, responsible use
Each workshop is approximately two hours, capped at 25 participants, and ends with a concrete working project — not "you understood X," but "you built X."
The six workshops
W1 — Build It
"An agent without memory is just a fancy autocomplete."
Build the agent from scratch: the model talks → streams → uses tools → runs the loop → remembers you across restarts via three plain Markdown files.
Deliverable: a streaming, tool-using, memory-keeping agent (CLI) with a capstone file you can point to and say "this is what I built today."
W2 — Empower It
"You're the architect — the LLM is your junior dev."
Give the agent real power: read/write files, run commands, and even write and run its own code — all behind a safety model you control.
Deliverable: an agent that acts on real files and self-corrects, safely.
W3 — Ship It
"If it doesn't run behind an endpoint, it's a toy."
Give the agent web tools and an email tool, then put it behind a FastAPI endpoint with a streaming client.
Deliverable: a locally served agent application with an HTTP API and a working client.
W4 — MCP (Advanced)
"Stop building tools. Start connecting."
Your agent becomes a client to any MCP server — borrowing tools it never had to write.
Deliverable: an agent powered by external MCP servers.
W5 — Skills (Advanced)
"A tool is an appliance; a skill is a recipe."
Package tools into reusable skills — SKILL.md recipes the agent loads on demand to do a whole job.
Deliverable: an agent with a skill library.
W6 — Loop Engineering + Production (Advanced)
"Stop prompting your agent — design the loop that prompts it."
Grow the W1 loop into a self-verifying goal-seeking engine with termination logic, budgets, evaluation, and ethics.
Deliverable: a production-pattern, self-driving agent with measurable quality.
Who this is for
- Basic Python required — functions, dictionaries, loops, terminal usage
- No previous AI-agent experience needed
- Works entirely with a local model (Ollama + qwen2.5) — no API key required to start
- "From scratch" refers to building without an agent framework — you still need basic Python experience
Register
Cohort size is limited to 25 participants. Pick a package below and click Register to reserve your spot.
Build a Real AI Agent — From Scratch · Ayman Hamed | AI Architect & Instructor · Simple but Impactful.