Who's building this
I got into IT in 2015, through an apprenticeship. Then iOS apps for brands every Austrian knows. Then almost four years running my own company, then corporate IT: projects, public tenders, handovers. On the side I taught myself AI and the dev skills that come with it, evening after evening, prototype after prototype. At some point the job no longer matched what I was building after hours. Early 2026 we parted on good terms. Today I build agent systems that keep running without me.
The tool
Everything here was built with session-orchestrator, my open-source plugin for Claude Code. It gives coding agents a process: understand, plan, build in waves, close out on evidence.
Open source, MIT.
/plugin install session-orchestrator@kanevryBuilt with it
I also build
FeedFoundry, internal: an Instagram content tool for my own brands. Text is rendered deterministically rather than generated; only the optional AI backgrounds pass a vision gate that rejects on doubt.
hermes-agent, a fork of the Nous Research agent: was my messaging hub, no longer developed.
ai-gateway, an internal LLM proxy: all agents run through one point, so costs and model choice live in one place.
sven, the third generation of my own agent infrastructure (first BIGG, then clank, now sven): works through mail, calendar, and notes every day, every action with an approval step.
A standalone community agent built on the same learnings: runs around the clock on its own server, answers semi-autonomously.
A B2B lead pipeline that merges 19 data sources across 3 servers every day.
A screening that finds SME succession candidates via managing-director age and digitization level.
A security layer for agent tool calls: prompt-injection detection with over 180 patterns, built in one hackathon weekend.
AI enablement in practice: weekly build sessions where people build with agents on their own projects. Not watching, doing.
How I work
Every session follows the same pattern: understand first, then plan, then the agents work through it in waves. A full build session runs a median of twelve agents; the peak was 117. I write the briefs, review the diffs, and approve when the result holds up in reality. Of the past 60 days, 58 looked exactly like that. Not an experiment anymore, everyday work. And I get teams working this way themselves: weekly build sessions, hands-on with their own projects.
“Today three agents ran in parallel. One builds, one tests, one documents, and I check in every two hours to see if the direction still holds. That's all it takes.”
| Role | Waves | Agents |
|---|---|---|
| Discovery | 57 | 151 |
| Impl-Core | 62 | 174 |
| Impl-Polish | 59 | 151 |
| Quality | 56 | 175 |
| Finalization | 55 | 30 |
| Ad-hoc roles | 26 | 29 |
Right now
- Venture studio at AI Factory Austria: turning signals into ventures.
- AI working group: weekly build sessions where 15+ people build with agents on their own projects.
- WalkAITalkie is on the App Store, AgenticCutter edits the videos about it on the side.
Where I have been
Four dates from the past six months.
If you're working on AI ventures, agent systems, or prototypes: get in touch.
What I'm not the right fit for: slides without a build, maintaining someone else's systems.