Venture 01 · Company Brain
Company Brain
Extract how work actually gets done, emit versioned skills, and fold every human correction back into a self-improving library.
The problem
Every company runs on know-how scattered across inboxes, Slack, tickets, and people's heads. Search-over-docs answers questions; it does not execute workflows. Agents need structured procedures — not another retrieval layer.
The loop
One source ingests resolved work. Extraction mines the actual procedure — not imagined best practice. A skill file is emitted in human-readable markdown with machine-parseable frontmatter. An agent drafts the resolution. A human approves or edits. The correction merges back into the skill. The diff log is the proof artifact.
“The skill files are the asset — not an internal cache.”
First wedge
Customer-support resolution: observable ground truth, text data, fast feedback loops. The acceptance bar is simple — an agent resolves a new, unseen case using only the extracted skill; after human correction, the next similar case measurably improves.
Why it compounds
Skill files are the asset — not an internal cache. A routing layer sends requests to the right departmental workflow. We dogfood inside ArcRos: building the brain and hardening our own operating model are the same effort.
Capabilities
- Procedure extraction from resolved cases
- Versioned skill files with diff history
- Multi-skill routing across workflows
- Human-in-the-loop correction merge
- Holdout evaluation and regression signals
- Research enrichment for skill maintenance
Dogfooded inside ArcRos
ArcRos is the first customer. Every correction our team makes merges back into the library — and the diff log of those corrections is how we prove the loop works.
Discuss a pilot