Agent Builders
Hands-on program for engineers building and shipping AI agents.
Contents
formatLive labs
Hands-on training for engineers who build AI agents. Over several days, teams design an agent, give it tools and memory, test it against real tasks and ship it with monitoring. Sessions are taught in English. Arclight engineers lead every lab. Nothing touches your production systems.
Labs cover Python and TypeScript, with any major model API.
Prerequisites
| Topic | Details |
|---|---|
| Who it suits | Engineers with working coding experience |
| What you learn | Agent design, tool use, evaluation and monitoring |
| Languages | Python or TypeScript |
| Bring | A laptop |
| Lab environment | A cloud sandbox; labs use a shared workspace |
| Delivery formats | On-site, online live, or blended over weeks |
Entry needs working coding experience. Cohorts are tailored to the stack your team uses, and we share the full syllabus on request.
Enrolment
Your engineers need a laptop with their usual development setup. We provide the sandbox tools used in the labs.
Share your stack and goals before the first session. Each agent you want to build must be agreed in advance. The programme adds only the labs and reviews you pick from the module list. Book a short scoping call so that the right engineers are in the room.
Bring a real use case to the first session. Labs build on it day by day. Your own data stays in your environment; sample data is used in shared labs. Lab notes are kept in a shared workspace.
Formats and access
| Module | Length | Outcome |
|---|---|---|
| Agent basics | Half a day | An introduction to agent basics for any engineer in your company |
| Build track | Agreed with you | Build a tool-using agent end to end; for engineers nominated by your sponsor |
| Production track | 2 days | Adds memory, approvals and monitoring to your agent |
| Evaluation track | Agreed with you | Tests the agent on realistic tasks, sets pass marks and wires up traces, so the team can ship with evidence |
Any engineer in your company can join the basics session. Only people nominated by your sponsor can take the build track. Participants keep their own lab notes and can edit or remove them.
Cancelled sessions leave the rest of the programme intact in all three delivery formats. If a session has to be rescheduled, we restore the whole timetable and keep every session linked to its module.
These arrangements are independent of your company's own access rules: this programme does not provide per-team or per-project isolation. If your organisation requires it, we agree lab access with you before participants work with real data.
Assessment
Each lab ends with a short assessment: the agent is tested on realistic tasks, pass marks are agreed in advance, and the team sees traces of what the agent did and why.
Module and exercise labels use plain language throughout. Relative dates follow the cohort's own calendar.
Production hand-over plan
Lab notes stay readable even when the programme is paused. Hand-over documents show each step and only the outcome.
Do not send customer data to shared labs. Other agents need the same careful treatment of data. This programme supplies no production code and no migration of your existing agents.
Lab environment
Labs default to a local sandbox seeded with practice data. Use these details only for practice. If you prefer a cloud sandbox, tell us in advance; hosted setups outside the labs need to be agreed with us. Both online and on-site modes are supported.
The default lab uses a local sandbox with no external service; the first run downloads the sample data. Cloud labs use a fresh disposable project.
The labs exercise tool use, memory, approvals, handoffs, failure handling, evaluation and monitoring. The late-failure recovery lab runs on cloud and on-premise setups and is skipped on the default local configuration, which has no rollback.
Starter code is used during the labs; shipping your own agent is a separate step. Build and test before handing over. No deployment is performed by these exercises.
Scope and limits
The shared workspace holds up to 200 notes per participant. Reopening a finished exercise restores its starting point. Repeated tasks or work split across steps may need manual correction, and finer progress tracking is outside this programme.
Contact
Questions about the programme? Write to contact@arclight.im.