Your engineers use agents.
Now make the work compound.
Six modules on your own codebase, ideally spread over six weeks. Fix a bug, sharpen a plan, build checks, and run a larger task twice. Each session leaves rules, context, and skills the next one can use.
See what your engineers buildSoftware engineers already using coding agents who want to take on harder work and judge the results.
One repository you ship to regularly, working tests, and backlog tasks ranging from a small bug to a larger feature or refactor.
Set up Claude Code, install the supplied skills, check repository and pull request access, and pick your first bug.
A correction is useful for the task in front of you. Save it as a rule, add the missing context, or build a check, and the next session can use it too. Your engineers practise this on their own codebase.
Modules 1–3. Short tasks with decisions you can inspect.
Fix a real bug in your own repository. Check what the agent actually read, write the failing test, review the fix, and take it through your pull request and ticket workflow. Capture what worked before you close the session.
Take back to work A tested fix and a personal rules file grounded in the session, ready for the next task.
Put a multi-file task through plan mode. Challenge the plan, work through unresolved decisions, and make the agent account for what it missed. Extract task-shaping rules from the planning session and your team’s own tickets.
Take back to work A reviewed plan and reusable rules for turning a vague request into work an agent can execute.
Work in parallel sessions on the same codebase. Review a feature’s access controls and threats, record a hardening decision, and author a test-strategy skill that fits your repository. Test the skill’s weak points before relying on it.
Take back to work A documented security decision and a test-strategy skill shaped by your code and testing conventions.
Modules 4–6. Two runs of the same task reveal what the system needs.
Scope a larger task, audit the context and rules you have built, and fill the worst gaps. Let the agent work without step-by-step steering. Keep the result as evidence for the next module, including where it stalled or went wrong.
Take back to work A first long-running experiment on real work, with a record you can inspect.
Find where the first run lost the goal, lost context, or produced plausible but wrong work. Build a check for the dominant failure. Give the agent a reference for the intended result and a plan, then run the same task again.
Take back to work A verification check, a reference, and a plan tested against a second run of the same task.
Compare the two runs. Decide which gaps need better memory, a stronger check, or a reusable skill. Remove a rule the evidence disproved. Look at how reliably the checks work across runs, then package a recurring task for reuse.
Take back to work Better rules and checks for your repository, plus instructions for turning a task you repeat into a reusable skill.
The exercises, handbook, and references are there when you need to do this again. The agents, workflows, and skills they build stay theirs. Training material usage is covered by your agreement.
I built the engineering and business training tracks at F-Secure, trained leaders personally, and developed internal trainers to carry the work forward. Today I’m AI Transformation Lead at WithSecure.
Around 500 people have been through my training programmes. I’ve personally trained about 100 of them.
We’ll agree what they need to learn, which modules to run, and how to fit the sessions around their work.
After the first group, we train people inside your company to teach the next groups.
Talk to Antti