Engineering notes · in English
Running a local model as an agent, honestly
What breaks when a 12B model has 21 tools and a real machine to act on, what the traces showed, and which guards actually held. Every number below was measured before it was written down.
2 notes · the rest of this site is in French
Guide · living document · source on GitHub
Building a local agent you can trust — what we learned, in order
The decisions behind Kora, a desktop agent on a 12B local model with real access to one person's files, mail and voice — each with the incident that forced it and the mechanism that holds it. Rendered from github.com/Kairos-commu/kora.
· 21 min
Catalogue · living document
Tool-calling failure modes on local models — a field catalogue
Every way a local model has broken tool calling in one real agent, September 2026, with the mechanical guard that stopped each one. Observed on gpt-oss:20b, qwen3.5:9b and gemma4:12b through Ollama.
· 12 min
92–93 %tool-choice rate, gemma4:12b with persona — the first fine-tune: 80 %
80 %multi-round scenarios, never inventing a capability, never re-asking after a refusal
~15 minone learning cycle: prepare → QLoRA → GGUF → ollama create → rate bench
Reference figures from spec/bench.md, github.com/Kairos-commu/kora