Buy the course

Hands-on course · In development

Build and maintain your own AI harness

Stop treating the AI assistant as a black box. Learn to design, build and run the layer of tools, context and workflows that turns a raw language model into a reliable working partner — and keep it working as models change.

Buy the courseSee the curriculum
  • 21modules
  • 5parts
  • 3layers
  • 1working agent stack

About the course

One project. Three layers.

  • L3

    Frontends & channels

    Slack, Mattermost, Telegram, Discord, e-mail, or your own web UI — with threads, identity and approval buttons.

  • L2

    Automation & events

    n8n and LangChain/LangGraph workflows triggered by webhooks, cron, Git and ticket events — with retries and human-in-the-loop.

  • L1

    Core agent: pi

    A reconfigurable agent with custom tools, extensions, skills and profiles, running headless behind your own agent gateway.

The idea

What is a harness?

The model predicts tokens. The harness is everything around it: the agent loop, tools, context management, model access, interfaces, safety controls, observability and extensibility. An agent is model + harness + a goal. Swap the model and you change how capable the agent is. Change the harness and you change what it can do, safely, and where.

  • Loop

    Prompt → model → tools → results → repeat.

  • Tools

    Files, shell, search, web, custom domain tools.

  • Context

    System prompts, AGENTS.md, memory, compaction.

  • Model access

    Providers, keys, fallbacks, cost control.

  • Interface

    CLI, editor, chat, automation and CI.

  • Safety

    Permissions, sandboxing, confirmations, audit.

  • Observability

    Logs, tracing, cost tracking, evals.

  • Extensibility

    Plugins, skills, hooks, MCP, sub-agents.

Build, don’t just watch

Every module ends with working code added to your own harness. You start with about 100 lines and add complexity only when it is needed.

Buy the courseTest mode: no real payment is taken.

What you learn

What you will be able to do

  • Explain what an agent harness is and how its parts fit together.
  • Implement a minimal agent loop with tool calling from scratch.
  • Configure and extend pi: custom tools, hooks, skills, packages and profiles.
  • Engineer context: system prompts, AGENTS.md, templates, compaction.
  • Use multiple model providers, including local models.
  • Run the agent headless behind an HTTP/WebSocket gateway.
  • Trigger and orchestrate agents with n8n and LangChain/LangGraph.
  • Put the agent in Slack, Mattermost and a custom web UI.
  • Secure, observe, evaluate and deploy the whole stack.
  • Maintain it over time: upgrades, versioning, avoiding prompt rot.

Curriculum

Curriculum

Built bottom-up: understand the harness, master the core agent, automate it, expose it to people, then run it in production. Draft outline — subject to change.

Part AFoundationsmodules 0–2
  1. Orientation & setupNode.js, Docker, pi installed, an AI provider connected — including how to create and connect a free agent.
  2. Anatomy of a harnessModel vs. harness vs. agent, the 3-layer stack, and a tour of existing harnesses.
  3. The minimal agent loopA ~100-line loop with tool calling, built from scratch.
Part BL1 · The pi agentmodules 3–9
  1. pi fundamentalsSessions and branching, models and providers, settings, AGENTS.md.
  2. Context engineeringSystem prompts, project context, prompt templates, compaction.
  3. ToolsBuilt-in tools, and designing custom ones: schemas, descriptions, errors, output size.
  4. Extensions & hooksTypeScript extensions, event hooks, permission gates, UI additions.
  5. Skills & packagesSKILL.md packs, sharing via packages, MCP through an adapter.
  6. Agent profilesOne pi, many agents: prompt + tools + model + skills per use case.
  7. Headless piPrint/JSON, RPC and SDK modes; build an agent gateway for the layers above.
Part CL2 · Automation & eventsmodules 10–13
  1. n8n essentialsSelf-hosted n8n, triggers, and calling the agent gateway.
  2. Event-driven agentsTrigger → enrich → agent → act → notify; retries, idempotency, queues.
  3. Human-in-the-loopApproval steps, wait-for-webhook, escalation.
  4. LangChain / LangGraphCode-first orchestration, and when to pick n8n, LangGraph or pi alone.
Part DL3 · Frontends & communicatorsmodules 14–16
  1. Slack botEvents API / Socket Mode, threads as sessions, slash commands, approval buttons.
  2. Mattermost & othersSelf-hosted option and a channel-adapter pattern for Telegram, Discord and Teams.
  3. Custom UIA minimal streaming web chat with tool-call visualization.
Part EProductionmodules 17–20
  1. Safety & permissionsChannel auth, sandboxing, secrets, prompt injection from chat input.
  2. Observability & evaluationTracing across layers, cost per user and channel, evals and regression tests.
  3. Deployment & maintenanceDocker Compose stack, upgrades, prompt drift, team sharing.
  4. CapstoneA personal or team 3-layer agent for a real workflow.

Who it is for

Who it's for

Software developers and technical professionals who already use AI assistants and want control over them — plus technical leads, researchers, power users and educators automating knowledge work.

Prerequisites

  • Comfortable with a terminal, git and one programming language.
  • Basic understanding of HTTP and JSON APIs.
  • Access to an AI provider — a free tier is enough. Module 0 shows you how to set one up.

No ML or deep-learning background needed.

How it works

How the course teaches

Build, don't just watch

Every module ends with working code added to your own harness.

Minimal first

Start with ~100 lines and add complexity only when it is motivated.

Vendor-neutral

Concepts over brands. Examples run against several providers.

Explain the why

Design trade-offs matter as much as the code.

Maintenance included

Upgrades and prompt drift are a first-class topic, not an appendix.

Real tasks

Exercises use realistic coding and knowledge-work scenarios.

Questions

FAQ

Do I need a paid AI plan?

No. A free tier is the minimum, and Module 0 walks you through creating a free account and connecting it.

Is the course tied to one vendor?

No. The concepts are vendor-neutral and the examples run against several providers, including local models. The core agent used throughout is pi.

What do I end up with?

A working, personally tailored 3-layer agent stack — core agent, automation and channels — and the knowledge to evolve it as models and APIs change.

When does it launch?

The course is still in development and the outline above is a draft. Launch details will be announced here.

Ready to build your own AI harness?

The course is still in development. Launch details will be announced here.

Buy the courseTest mode: no real payment is taken.