Operator Workspace Blueprint: giving humans and agents one source of truth
How to structure a knowledge system and agent stack so your operating context compounds instead of scattering across chats and docs.
who it's for — Operators and teams who've outgrown ad-hoc AI use and want the system that runs underneath.
Most teams adopt AI one chat at a time. Context lives in scattered histories and people's heads, so every agent starts cold and nothing compounds.
An Operator Workspace fixes the substrate: a single source of truth that both humans and agents read from and write to. Get this right and every session makes the next one smarter.
Start with a structured, plain-text knowledge base (we use an Obsidian/Markdown vault). It holds three things: SOPs (how work is done), decisions (what you chose and why), and project context.
Plain text matters — it's diffable, version-controlled, and equally readable by a person and an agent. Avoid locking your operating knowledge inside a SaaS you can't query.
Give each agent one clear job — research, drafting, review, operations — rather than one do-everything assistant. Specialization makes behavior predictable and easy to trust.
Every agent reads from and writes back to the vault. Outputs aren't trapped in a chat window; they land in the source of truth where the next agent and the next person can use them.
The workspace earns its keep when work flows through it: a decision gets recorded, an agent drafts against it, a human reviews, and the result goes back into the vault.
Onboarding becomes reading the workspace instead of shadowing someone for weeks, because the operation is written down in a form both humans and agents can use.
Don't dump raw transcripts into the vault — curate durable facts and decisions, or you bury signal in noise.
Don't build one giant agent. Don't store operating knowledge somewhere agents can't read. And don't expect a workspace tuned to one team to transplant to another without customization.
- ▸Operating knowledge lives in plain text (diffable, version-controlled)
- ▸Vault holds SOPs, decisions, and project context as distinct layers
- ▸Each agent has one clear job, not a do-everything mandate
- ▸Agents read from and write back to the same vault
- ▸Work flows through the loop: record → draft → review → store
- ▸Durable facts are curated; raw noise is kept out
- ▸Onboarding is reading the workspace, not shadowing a person
Do I need Obsidian specifically?
No — the principle is plain-text, version-controlled knowledge that humans and agents can both read. Obsidian is what the studio uses, but the blueprint works with any Markdown-and-Git setup.
How many agents should I start with?
Fewer than you think. Start with two or three with clear jobs (say research, drafting, review), prove the loop works, then add specialists as real needs appear.