# AI agent development workshop for engineering teams

Source: https://tpiros.dev/ai-engineering-workshop/building-agents/

A hands-on workshop in building AI agents, for your engineering team. Your engineers build AI agents in TypeScript or JavaScript, writing the moving parts themselves as we go: agent patterns, tool use, harness engineering and evals, against OpenAI, Anthropic or Google, whichever you build with. The syllabus is my free book, AI Engineering for Web Developers, and the call picks the chapters. They need the basics of one of those two languages and of large language models, and no machine learning background. For one engineering team. This is one half of the AI Engineering Workshop at https://tpiros.dev/ai-engineering-workshop/, and the call can add the other. Remote or on-site. A paid engagement, quoted against scope; nothing here is priced.

The written form of this half is the free book AI Engineering for Web Developers at https://tpiros.dev/ai-engineering-for-web-developers/: read online, no sign-up. Building agents draws on its chapters.

## What the building-agents half covers

The building-agents half of the AI Engineering Workshop has an engineering team write the parts of an AI agent in TypeScript or JavaScript, topic by topic from the free book AI Engineering for Web Developers. These are the topics the call chooses from. Each is a chapter of the book, so your engineers can read it before or after.

- **Tool use.** Defining a tool, calling it, and steering the model's choice when it has several. ([Tool use](https://tpiros.dev/ai-engineering-for-web-developers/tool-use/))
- **Structured output.** Model output that parses against a Zod schema, and the schema patterns that bite. ([Structured output](https://tpiros.dev/ai-engineering-for-web-developers/structured-output/))
- **Agent patterns.** From a single agent to the coordinator and agent-as-tool, split by who picks the next step: the workflow or the agent. ([Agent patterns](https://tpiros.dev/ai-engineering-for-web-developers/agent-patterns/))
- **Harness engineering.** The runtime split into brain, hands, session and feedback behind narrow contracts. ([Harness engineering](https://tpiros.dev/ai-engineering-for-web-developers/harness-engineering/))
- **Memory.** Working, conversation and long-term memory, with extraction, recall and forgetting. ([Memory](https://tpiros.dev/ai-engineering-for-web-developers/memory/))
- **Multi-agent handoff.** A typed handoff payload and result between agents, and A2A when the work crosses a process. ([Multi-agent and handoff](https://tpiros.dev/ai-engineering-for-web-developers/multi-agent-and-handoff/))
- **Evals.** A three-layer eval suite, a starter eval set, and an LLM judge kept honest. ([Evals](https://tpiros.dev/ai-engineering-for-web-developers/evals/))
- **Guardrails.** Input and output filters, tool authorisation and an audit log, against prompt injection and bad actions. ([Safety, guardrails, and prompt injection](https://tpiros.dev/ai-engineering-for-web-developers/safety-guardrails-and-prompt-injection/))
- **Sandboxing.** Running agent-generated code in isolation, and deciding which tools need it. ()
- **Observability and cost.** Per-run logs and traces, per-user budgets, and streaming the result to the user. ([Observability, cost, and deploy](https://tpiros.dev/ai-engineering-for-web-developers/observability-cost-and-deploy/))

**What the team writes.** Your engineers type the code themselves, against OpenAI, Anthropic or Google, on content shaped around what they are building. Depending on the topics the call picks, that can be a tool definition and the loop that calls it, a schema the output has to pass, a harness with its layers split apart, or an eval set with a judge. They leave with the code they wrote, and the chapter behind each topic stays free to read.

## How it works

Your engineers leave with the code they wrote. I explain each moving part and they write it, on content shaped around what they are building. No pre-work for attendees. Remote or on-site.

1. A 30-minute call. What your team is building, which coding agents they use, and which half of the workshop they need. Often it is both.
2. Shaped around your team. I put the content together for your team and the work in front of them. How long it runs depends on the scope we agree.
3. The workshop. Remote or on-site. I explain each part, and your engineers do it as we go, on their own code with their own coding agent or by writing the agent themselves.

## Two halves, one workshop

The call decides which half your team needs, or both, and in what order.

- **Working with coding agents**: Cursor, Claude Code, Codex or Antigravity, whichever your team uses, in any language, usually on your own codebase. Giving the agent the context it needs with AGENTS.md, skills and rules files, planning before it writes, reviewing what it wrote, and deciding which work to hand it in the first place.
- **Building agents**: Agents, agentic workflows, harness engineering and the other moving parts, with your engineers writing the code as we go. The syllabus is the free book, and the team needs basic JavaScript or TypeScript.

Each half has its own page: working with coding agents at https://tpiros.dev/ai-engineering-workshop/coding-agents/, building agents at https://tpiros.dev/ai-engineering-workshop/building-agents/.

## Who is in the room

One engineering team. For working with coding agents, any language. For building agents, basic JavaScript or TypeScript and a basic understanding of large language models. No machine learning background for either. Where the wider problem is that most of an organisation never got past asking these tools questions, that's different work: https://tpiros.dev/power-user-gap/.

## What they leave able to do

Depending on what the call settles, they hand real work to the coding agent they already have and know how to check what comes back, or they understand how agents are built because they wrote the moving parts themselves. Often both.

## Before you book

**Do we need an agent framework?** No. The team writes the moving parts in plain TypeScript or JavaScript first, so they know what a framework would be doing for them. The book also rebuilds its assistant on Google's ADK and on Vercel's eve, if the call wants a framework in.

**Our team writes Python. Does this half work for us?** This half is written in JavaScript or TypeScript, the language of the book, so the team needs the basics of one of them. The coding-agents half works in any language.

**Does the team need a machine learning background?** No. Basic web development in JavaScript or TypeScript and a basic understanding of large language models are enough.

**Does the team have to read the book first?** No. There is no pre-work. Each topic in the workshop has its chapter, free to read before or after, so the team can go back to it.

**What do attendees need to prepare?** Nothing. There is no pre-work. The call decides what goes in.

**What does it cost?** It is quoted against the scope we agree on the call. There are no individual seats and no published prices.

Questions about model providers, recording and the teaching itself are answered at https://tpiros.dev/ai-engineering-workshop/.

## Next step

Book a 30-minute call at https://calendly.com/tamas-piros/30min, or email hello@tamaspiros.com with what your team is building and what they need to learn. It's one quote, for one engineering team and the scope we agree on the call, one half or both, with no individual seats, so the figure you forward is the whole workshop.