Berlin | Claude Workshop
Build agents that debug and review with Claude Code. A hands-on evening: bring a laptop, leave with a working agent and a feature built by a crew of Claude Code sessions.
- When: Thursday 22 October 2026, 18:00 to 21:30
- Where: P.19 Studio, Prinzessinnenstr. 19, 10969 Berlin
- Hosts: Mehul Patel and Markos Sfikas, Claude Community. GlassFlow hosts the evening at its venue.
- Event page: luma.com/claude-fxrb
- Filming: the evening is filmed. If you prefer not to appear on camera, tell us at the door.
Do the setup at home before you come. The venue Wi-Fi is shared by 65 people, so installs and downloads on the night are slow.
Tares is open source. Both parts tonight run on it. If you like what you build, star the repo on GitHub: it helps others find it.
★ Star Tares on GitHub
The two parts
Part 1: Build an AI SRE with a handover agent
You build two agents that watch a small live service, break the service on purpose, and watch the agents find the root cause.
- You watch a demo service with steady traffic: its Prometheus metrics, its error logs and the alerts Prometheus fires. Tares puts all of it on one timeline for the service.
- You build a watcher on a small, cheap model that looks every 2 minutes, and a root-cause agent on a bigger model that only runs when the watcher says “investigate”.
- You see a finding appear 2 to 4 minutes after you break the service: what is failing, since when, the likely cause and a next step. Then you ask Claude Code the same question and compare.
- You take home the watch-then-escalate pattern, and how to point it at your own services.
About 45 minutes. Start Part 1.
Part 2: Software factory
You set up a software factory on your laptop: a crew of Claude Code sessions that turns a plan into pull requests merged on main. Then you give it a small test job to watch the whole loop run once.
- You build the crew: an orchestrator, a builder and a reviewer, running in the background on your own Claude Code plan, with Tares keeping the plan, the tickets, the reviews and your answers.
- You test it on a toy job: a click counter for a tiny link shortener, your own copy of a template repo with tests and CI. You plan it in a short interview with Claude: two tickets.
- You watch the orchestrator hand the plan to the builder, which opens one pull request per ticket. The reviewer reviews each one, and once CI is green the builder merges it.
- You take home a factory that works on any repo: point it at a real project next.
About 30 minutes. Start Part 2.
Agenda
| Time | What |
|---|---|
| 18:00 | Doors, food, drinks, and the setup help desk |
| 18:15 | Welcome and Claude Community intro |
| 18:30 | Tares in 10 minutes |
| 18:40 | Part 1: build an AI SRE |
| 19:45 | Part 2: software factory |
| 20:15 | Short break |
| 20:30 | Building continues, and stretch goals |
| 21:00 | Show and tell |
| 21:20 | Wrap-up and what’s next |
| 21:30 | Networking |
Local or cloud
You can run each part against a Tares on your laptop (local) or against a Tares workspace we host for you (cloud).
- Local: everything runs on your laptop. Part 1 local needs Docker and your own Anthropic API key.
- Cloud: a free Tares workspace with $10 of model credit. You get the join link on the night. No API key needed.
Both are fine. If Docker or the Wi-Fi gives you trouble on the night, switch to cloud. The guides here are the local versions; the cloud versions follow before the workshop.
Downloads
Everything the two parts use, in one place.
| File | For | What it is |
|---|---|---|
| tares-demo.zip | Part 1 | The demo stack: docker-compose.yml and inject.sh. Unzip it, then run docker compose pull inside: the images (about 600 MB) still come from the internet. |
| factory-workshop-starter | Part 2 | The link shortener you copy with Use this template. |
| click-counter-spec.txt | Part 2 | The ready-made spec, to skip the planning interview. |
| workshop-slides.pdf | Both | The slides for Part 1 and Part 2. Also on the Slides page. |
Keep going
After the workshop, point what you built at your own systems.
Part 1: AI SRE
- Build an AI SRE : the same demo, then how to swap in your own metrics, logs, alerts and deploys.
- Watch, then escalate : the watcher and root-cause pattern from Part 1, for any sources.
- Connectors : Prometheus, Alertmanager, Docker logs, OpenTelemetry, GitHub, Postgres, webhooks and more.
- Tares cookbooks : a complete incident-response agent in code, if you would rather write the agent than configure it.
Part 2: Software factory
- Get started with the software factory : run a crew on your own repos.
- A project end to end : one real project followed from the spec to eight merged pull requests and a deploy.
- Autonomous Claude Code sessions : how the crew’s sessions keep working in the background while you are away.
Tares and Claude Code
- Tares docs and the Tares repo on GitHub . A star helps others find it.
- Connect your agents to Tares over MCP.
- Claude Code docs .