F-01One day09:30 to 17:00

Agents on Prod

As soon as an agent is allowed to touch more than a single file, its behaviour changes. It finds another way when the described one does not work. It reports success for things it never ran. And sometimes it fixes a problem nobody had along the way, so the real bug can no longer be reproduced afterwards. This day is about the gap between the impressive demo and a workflow you can rely on.

Reserve a seat

Next date Thursday, 12 November 2026, €690 net as pilot price.

Duration

One day, 09:30 to 17:00

Group

4 to 8 people

Price

€690 net, regular €890

Venue

Grell37 in Berlin Prenzlauer Berg or at your place

Who it is for

Developers, tech leads and freelancers who already use a coding agent and now want to let it touch code that somebody is liable for.

What you bring

  • Your own laptop with Claude Code or a comparable agent up and running
  • A real repo you are allowed to work on. No sample project
  • Basic Git routine

What you take away

  • A CLAUDE.md for your own repo, written and tested
  • A slash command and a subagent of your own that still work on Monday
  • A rule set for the team: which actions the agent never takes on its own
  • A checklist for the moment the agent claims to be done

Not on the programme

  • Prompt engineering for its own sake
  • Model comparisons and benchmarks
  • Building your own agent frameworks

Agenda

The times are a plan. If a topic is burning in the group, the rest shifts.

09:30

Arrival and setup check

Everyone's agent is running, everyone has their repo with them. Whoever is still installing in the morning loses half of it.

10:00

How an agent actually works

Context window, tools, permission modes, what gets lost when the context is compacted. Live on a real repo, no slides. By the end you know why the agent starts making things up after two hours.

11:00

Break

11:15

Preparing the repo

What belongs in CLAUDE.md, what in a knowledge base, what in the agent's memory and what nowhere at all. Without that separation the agent starts from zero every morning. You write the file for your own repo and try it on a real task straight away.

hands-on

12:30

Lunch

13:30

The ten rules

Every rule with the incident behind it, with a date and with what it cost. Credentials in the transcript. A firewall rule that takes the production API offline for two hours. A cache that halves a media library. It is the most uncomfortable block of the day.

14:45

Extending the agent

Slash commands, subagents, skills and a small MCP server on your own API. You build a command and a subagent that you keep afterwards.

hands-on

15:45

Break

16:00

When it goes wrong

The recurring failure patterns: the fix that only repairs the surface. The detour when the named path does not work. The rebuild where one line would have done. And the classic that costs the most hours: the agent reports done without having run it. We practise spotting that in the wording before you notice it in production.

16:30

Rollout in the team

Cost and consumption, model choice, branch rules, protection on main, a guard in CI and the short list of things the agent never does without a human.

16:50

Wrap-up

Everyone leaves with three rules to put into practice on Monday. Written down. Experience says a resolution lasts until Tuesday.

Dates

A date runs from four registrations. Until it is confirmed, your registration is non-binding and free. You will know at least ten days in advance whether the date goes ahead. The invoice comes only after that.

09:30 to 17:00

Thursday, 12 November 2026

Grell37, Grellstraße 37, Berlin Prenzlauer Berg. Greifswalder Straße S-Bahn station, four minutes on foot.

€690

Pilot price

09:30 to 17:00

Thursday, 10 December 2026

Grell37, Grellstraße 37, Berlin Prenzlauer Berg. Greifswalder Straße S-Bahn station, four minutes on foot.

€890