A two-day AI pair programming bootcamp works best when day one builds individual skills (agent setup, context files, task briefs, test-driven verification) and day two applies them to a team capstone project with parallel agents, code review and a retrospective. Every block should be at least half hands-on, and the bootcamp should end with conventions the team commits to its repository.
On this page
A single day of agentic coding training is enough to learn the basics. Two days allow something more valuable: a realistic team project in which participants experience the full cycle of planning, delegating, reviewing and integrating agent-written code under time pressure. This blueprint describes a two-day AI pair programming bootcamp that we have refined in practice. Adapt timings and tools to your team, but keep the structure: skills first, then integration under realistic conditions.
Design principles
Three principles shape the curriculum. First, hands-on beats explanation: no block should be more than half presentation. Second, verification is part of every exercise: no task counts as done without tests or a reproducible check. Third, the team owns the output: context files, prompts and rules created during the bootcamp go into the team's repository, not into a trainer's slide deck. The tool can be Claude Code, OpenAI Codex, Cursor or a combination; the structure works for all of them.
Day one: individual skills
| Time | Block | Outcome |
|---|---|---|
| 09:00 | How agents work: the loop, tools, permissions, limits | Shared mental model and working setup on every laptop |
| 10:00 | Codebase exploration with an agent | Each participant explains an unfamiliar module, verified against the code |
| 11:00 | Context files: CLAUDE.md, AGENTS.md or project rules | First committed context file for the bootcamp repository |
| 13:00 | Writing task briefs | Three backlog tickets rewritten as agent briefs with acceptance criteria |
| 14:00 | Test-first delegation | A bug fixed with a failing test written before the fix |
| 15:30 | Recovering from agent failures | Practice with stuck loops, wrong assumptions and rollbacks |
| 16:30 | Day one retrospective | List of open questions and friction points for day two |
The context file block deserves extra time when the team has never written one. A good exercise is to write the file, give the agent a task that depends on it and then remove one section to see whether the agent behaviour changes. Sections that make no difference are cut; mistakes the agent still makes become new, more specific instructions. The recovery block is often underestimated: participants deliberately provoke failures, such as ambiguous instructions or a missing dependency, and practise noticing them early instead of letting the agent spin.
Day two: the capstone project
Day two runs as a compressed sprint. The trainer prepares a small but realistic product increment, for example a new API endpoint with a matching UI screen, an export feature or an integration with a mock external service. The team plans it together, splits it into agent-sized tasks and delivers it by late afternoon.
- Planning (09:00). The team writes a short spec and splits it into four to eight tasks with clear interfaces between them.
- Parallel delegation (10:00). Pairs work on separate tasks, each with an agent, using branches or separate worktrees to avoid conflicts.
- Mid-day integration (12:00). First merge round, resolving interface mismatches and conflicts.
- Review drill (13:30). Each pair reviews another pair's pull request, supported by an agent, using a shared checklist.
- Hardening (14:30). Edge cases, error handling and missing tests, with agents writing additional tests against the spec.
- Demo and retrospective (16:00). The increment is demonstrated, and the team agrees on conventions to keep.
The parallel delegation phase surfaces the real coordination challenges of agentic work, which we discuss further in multi-agent development workflows. The review drill builds on exercises from our AI-assisted code review workshop.
Assessment without exams
Bootcamps for professionals should not end with a written test, but the trainer and team lead benefit from a light assessment. Useful criteria are whether each participant merged at least one reviewed agent-assisted change, whether the capstone increment passes its tests, how many review comments pointed to real defects and whether the retrospective produced concrete, written conventions. These signals also serve as a baseline for later measurement of training impact.
Preparation for the organiser
Two weeks ahead, confirm tool licences and accounts, prepare the bootcamp repository with a working build and test setup and check network access from the training room. Send a short survey on prior experience so the trainer can form balanced pairs. One week ahead, share pre-reading and ask participants to install the tools. On the day, make sure a tech lead attends both days; without one, the conventions agreed on day two lack authority.
Common adjustments
Teams new to agents may need more time on day one and a smaller capstone. Experienced teams can replace the exploration block with advanced topics like hooks, MCP integrations or running agents in CI. Mobile teams can make the capstone a feature in an iOS or Android app, including a build for internal testing. Data and backend teams can choose a pipeline change with schema migration and validation checks, which exercises the same skills of decomposition, verification and review in a different domain. For a one-day version, our workshop agenda template condenses the essentials.
Running the bootcamp with us
We build and launch digital products with this exact way of working, and our trainers are practitioners. The bootcamps and workshops at neooutput run remote or on-site, in English or German, for Claude Code, Codex and Cursor.
Frequently asked questions
What group size suits a two-day AI pair programming bootcamp?
Groups of roughly six to twelve participants work well, because the capstone needs several pairs but the trainer must still support each one. Larger groups need a co-trainer.
Should participants pair with each other or with the AI?
Both. Human pairs working with one agent share techniques and catch each other's blind spots, which accelerates learning compared to everyone working alone.
Can the capstone use our real product backlog?
Yes, if the tasks are small enough to finish in a day and the repository is cleared for use with the chosen tool. Otherwise a realistic sample increment avoids pressure from production deadlines.