Acasuso.
← All work

A second brain with an AI executive team

Built Stratas Brain: the company's knowledge in one git-backed Obsidian vault, with scheduled agent loops that write a sales pack every weekday, a growth brief and a Monday board memo, delivered by email and Telegram.

My contribution

Designed the vault structure and registry, implemented the scheduled agent loops and delivery by email and Telegram, and wrote the runbooks and kill switches that keep it safe to run unattended.

Stratas Brain as a graph, drawn from the real vault: 472 notes and 750 links grouped by folder, with the most-linked notes in gold. Note names withheld.Open full-size image ↗
Stratas Brain as a graph, drawn from the real vault: 472 notes and 750 links grouped by folder, with the most-linked notes in gold. Note names withheld.

Situation

A two-founder company has no chief of staff, but the work a chief of staff does still has to happen: collecting signals, keeping one version of the truth, and turning both into weekly decisions. Company knowledge sat in chats, notes and inboxes, and the weekly sales, growth and board thinking happened only when someone found the time.

Task

Keep one source of truth for every company, person and product, and make the recurring management work happen on schedule without adding meetings or another tool to check.

Action

Architecture

Plain markdown in git. The vault is an Obsidian folder in a git repository, synced every few minutes. A master registry indexes every company, person and product, and the rules for agents and people live in one file at the root.

Loops as scheduled agents. GitHub Actions run the loops: signal rollups, a sales pack Monday to Saturday, a growth brief, and a board memo every Monday. Each loop writes its output back into the vault, so the history is versioned with everything else.

Delivery where people already are. Briefs go out by email and Telegram. A Telegram channel answers questions from the vault, captures notes and can trigger a loop on demand.

Off switches for everything. Every loop has a one-line kill switch and a runbook, so an agent that misbehaves is stopped in seconds, not debugged in a panic.

Implementation

  • Vault structure and registry, with folders for signals, packs, briefs and operations.
  • Four scheduled loops with email and Telegram delivery.
  • A Telegram channel running on Claude Code for questions, notes and triggers.
  • Runbooks for each loop.

Verification

  • Each loop writes its output into the vault, so every brief is versioned and reviewable in git.
  • Every loop has a kill switch and a runbook.

Result

  • Recurring management work happens on schedule: the sales pack is waiting every weekday morning and the board memo every Monday.
  • One place for every fact about the business, readable by people and agents alike.
  • History for free: every brief and every change is in git.

Skills evidenced

  1. Agent loops on scheduled infrastructure with delivery by email and chat.
  2. Knowledge management in plain text designed for both people and agents.
  3. Operational safety: kill switches and runbooks for every automated loop.

Portable lessons

  1. Plain markdown outlives every tool around it. Agents read it, people read it, and git keeps the history.
  2. Write agent output back to the source of truth. A brief that only lives in an inbox is lost by Friday.
  3. Every loop gets an off switch on day one.

Want to discuss the decisions behind it?

Over fifteen years of consulting and growth, and hands-on AI delivery from discovery to daily operations. I can share my CV and talk through how I connect business outcomes with systems that run.

Available for a chat. In person in Stavanger, or remotely worldwide.