The story
- A new door: OutSystems O11 released an early-access MCP. Some tools were missing, and some calls stopped halfway.
- Hit a wall: the first builds failed to publish.
- Go and look: on each error the agent searched the OutSystems library, the official samples and modules I had built before.
- Fix and redeploy: it applied what it found and published again, until it went green.
- Remember it: every fix was saved as a memory note and a reusable skill.
- Day 2, faster: the same problems were now solved on the first try, and the loop became stable.
- Tell the makers: the gaps, with steps to reproduce and suggestions, went back to the OutSystems product team.
What was delivered
- An agentic loop on PI console with the private local LLM, building on O11 through the early-access MCP.
- Error-driven self-learning: library, samples and past work as the places to look, and memory plus skills as the place to keep what worked.
- Two days of training that turned a shaky early-access connection into a stable way of building.
- Written feedback for the OutSystems product team on the MCP's current limits.
Why it matters
Early-access tools rarely work perfectly. A loop that learns from its own errors makes them usable now, and everything it learns stays: the next build starts from the skills the last one earned. And because it runs on the local LLM, the code and the know-how never leave the building.