A company run on AI agents
Acorn Labs runs on a team of AI department agents in one workspace: a manager agent hands out bounded jobs, departments research, draft and check, and nothing is sent, spent or published without my yes. Every decision and mistake goes into one written record.
Training and exams for AI agents
Each department gets a study pack for its real job and a scenario exam. A separate model marks the answers, flags anything copied from the model answers, and the results are kept. A trainer writes the next course from each department's weak spots.
Search that rates its own sources
A research tool that splits an answer into claims, re-opens every source to check the quote is really there, and rates each source: official, news, reference or weak, with the page date. A correct quote from a weak or AI-farm page is marked as weak.
Numbers that can say "I don't know"
A usage dashboard for the agent team that shows when a figure can't be read, instead of making one up. It took four passes to get there, and the write-up says why.
Holdfast (free)
One shared memory and rulebook for Claude, Grok, ChatGPT and coding agents: a Python MCP server with a web dashboard, setup guide and undo history. Two testers. It's free. See Holdfast →
A marketing memory (growth-os)
The system behind the marketing: what buyers say in their own words with the source, content made from it, warm follow-ups, tests, pages built to be cited by AI assistants, and a weekly one-page cockpit. How it works →