# The Agent Workshop

Source: https://tpiros.dev/agent-workshop

A hands-on half-day for an engineering team on building agents that work more than once, delivered remotely or on-site, extending to multi-day where a team wants to apply it to their own systems. A paid engagement, quoted against scope; nothing here is priced. There is no free developer session: the free book at /ai-engineering-for-web-developers and the technical articles are the way to judge the teaching first.

## The premise

Agents do not fail because the prompt was not clever enough. They fail because a tool returned something the model could not interpret, or the context filled up and the agent lost the thread, or nothing in the system could tell a good run from a bad one. All three are engineering problems with known solutions.

## What you build

- The harness: what the agent can see, what it may do, what happens when a tool fails, and how the loop knows to stop.
- Tool contracts: tool descriptions as interface design, useful error messages, return values shaped so the model does not have to guess.
- Context strategy: where context goes, why long-running agents degrade, and the practical answers (compaction, handoff, scratch files, sub-agents).
- Evaluation: a small evaluation over your own agent, broken deliberately, so you can watch it catch the failure.

## Prerequisites

JavaScript, TypeScript or Python; at least one prior LLM API call; a machine you can install packages on and a key for any major provider. No machine learning background, no particular framework, and the exercises cost a few cents of model usage to run.

## Related

The same material runs inside companies as part of the AI Engineering Programme at https://tpiros.dev/programmes. The free book at https://tpiros.dev/ai-engineering-for-web-developers covers the ground before it.

## Next step

Details are at https://tpiros.dev/agent-workshop. For the long form, ask about the AI Engineering Programme by emailing hello@tamaspiros.com.