A reusable one-day agentic coding workshop agenda has five parts: setup and mental model, context files, task briefs with test-first delegation, a realistic team exercise with review, and a closing session that commits team conventions. Each block should be at least half hands-on and produce an artefact that stays in the repository. The template below works for Claude Code, OpenAI Codex and Cursor.
On this page
Organising an internal agentic coding workshop is easier with a proven structure. The agenda below is the one-day format we adapt most often for development teams. It works with Claude Code, OpenAI Codex, Cursor or a mix, and it is designed so that every block produces something the team keeps: a configuration, a context file, a set of briefs, merged code or written conventions. Use it as a starting point and adjust timings to your group.
The one-day agenda
| Time | Block | Format | Artefact produced |
|---|---|---|---|
| 09:00 to 09:45 | Kick-off: goals, mental model of the agent loop, permissions | Short talk and live demo | Agreed workshop goals |
| 09:45 to 10:30 | Setup check and first task: explain an unfamiliar module | Individual, hands-on | Working setup for everyone |
| 10:45 to 12:00 | Context files: write and test CLAUDE.md, AGENTS.md or rules | Small groups | Committed context file |
| 13:00 to 14:15 | Task briefs and test-first delegation on real tickets | Pairs | Briefs and first agent-assisted pull requests |
| 14:30 to 15:45 | Team exercise: small feature split across pairs, then cross-review | Pairs and group review | Merged, reviewed feature |
| 16:00 to 16:45 | Security guardrails and failure recovery | Demo and discussion | Permission profile and review checklist |
| 16:45 to 17:15 | Conventions and next steps | Facilitated discussion | Written team conventions and owners |
Breaks are deliberately generous. Agent work is mentally demanding in a different way than coding: participants constantly evaluate output, and attention drops faster than in a typical training day.
Notes on each block
Kick-off
Keep slides to a minimum. The demo should include a failure, for example the agent misreading a requirement, so expectations are realistic from the start. Explain how the tool's permission settings work and which profile the team will use today.
Setup and first task
Asking the agent to explain an unfamiliar part of the codebase is the ideal first task: low risk, immediately useful and a good test of whether tools, accounts and repository access work. Have a helper ready for setup problems.
Context files
Groups draft the file, then test it with a task that depends on its content. If the team uses several tools, decide on one source of truth. Our guide on CLAUDE.md and AGENTS.md best practices can serve as pre-reading.
Briefs and test-first delegation
Each pair picks two or three real tickets, rewrites them as agent briefs with acceptance criteria and asks the agent to write a failing test before implementing the change. This block usually produces the biggest shift in how participants work. Ask pairs to keep their weakest brief alongside the improved version; comparing the two in the plenary makes the effect of acceptance criteria and constraints visible to everyone, and the pairs can reuse the better briefs as templates for their own backlog after the workshop.
Team exercise and review
A small feature is split into parts with clear interfaces, so pairs can work in parallel. The cross-review uses the team's draft checklist and reveals differences in quality expectations that are worth discussing openly.
Security and conventions
The security block covers secrets, prompt injection, dependency verification and MCP servers. The closing session turns the day's observations into a short written list of conventions, each with an owner.
Preparation checklist for organisers
Most workshop problems can be prevented in the two weeks before the date:
- Confirm tool licences, accounts and data approvals for every participant.
- Prepare a repository with a working build and test setup, either your own or a realistic sample.
- Collect five to ten real tickets of suitable size for the brief exercises.
- Check network access, proxies and firewalls from the workshop location or home setups.
- Send a short experience survey and form balanced pairs.
- Book a follow-up session four to six weeks later.
Variations for different teams
For teams new to agents, extend the setup and context file blocks and shrink the team exercise. For experienced teams, replace the first task with advanced topics such as hooks, subagents, background agents or CI integration. For mixed audiences with product managers, add a prototyping block using the same tools. For a two-day version with a full capstone project, see our two-day AI pair programming bootcamp. For a remote delivery, split the agenda into three half-day sessions across one or two weeks, keeping the same order of blocks; our comparison of on-site and remote workshop formats covers the trade-offs.
Measuring whether it worked
Define success before the workshop: for example, that every participant merges at least one reviewed agent-assisted change within two weeks, that the context file is updated at least once in the first month and that review rework rounds on agent-assisted pull requests decrease over time. These are simple to observe and say more than a satisfaction survey.
Facilitation support
Running the agenda internally works well if someone on the team already has deep practical experience with the tool. If not, an external facilitator saves preparation time and brings tested exercises. We build and launch digital products with this workflow daily and run the agenda for teams through agentic coding workshops at neooutput, remote or on-site, in English or German.
Frequently asked questions
How many participants suit this one-day agenda?
It works well for roughly six to twelve participants with one facilitator. Larger groups need a second facilitator, mainly for setup help and the pair exercises.
Can the agenda be used for Cursor instead of a terminal agent?
Yes. Replace the context file block with Cursor project rules and adapt the demos to the editor; the remaining blocks work unchanged.
What if participants cannot use the company's real code?
Use a realistic sample project in the same language and framework, ideally with some legacy complexity, and prepare tickets that resemble your actual backlog.