No hype, no proof-of-concept that disappears in a drawer. This series documents a complete AI agent operating system in production at JUNE, a German legal-tech company, written by Chris, who built it there and is also part of our team, not as a bumbleflies client project. Autonomous bots that open pull requests while the team sleeps. And what we learned along the way: a night that burned tokens in an idle loop, a read email that sat for days, an endless self-restart loop. Every guardrail here is an experience that became a rule.
Several agent pillars on shared foundations. Autonomous agents that open pull requests, a skill marketplace for company knowledge, a cockpit that plans the day. This is what AI looks like when it doesn't end in a demo.
The project management tool is persistent state and the trigger for automation. The team chat is the second foundation: interaction and transport. Agents coordinate through persistent artifacts, not direct calls.
A missing column in an export travels from support chat into a ticket, through enrichment, testing, and batch rollout to regression and docs. One concrete trace, no internal details.
Many workflows, a lot of processing steps, no human in the loop. A self-learning ticket router and a privacy filter that lets the model write, but doesn't believe a single word.
Many plugins, even more skills, activated by natural language. The same code runs for the human on a laptop, the agent in a container, and the CI. How to version company knowledge instead of copying it.
A word in the chat wakes an agent. It implements, opens a PR, reports back. Several personas from one building kit. And the night an agent burned a lot of tokens in an idle loop.
The personal side of the stack. A meta-agent that scans many sources, reconciles every loose thread against tickets, and makes every task one message away from being done.
We do. Send us two sentences about your topic. We'll reply with an honest assessment.
Book a first conversation