BusyBeaver
agent · you talk to it
Takes requests from anyone in the room, always within that person's permissions, and calls other agents when a job needs them.
acts after a yes
Invite your team, set roles, bring your own model. Every action is authorized and attributable.
# ops · 3 people · 2 agents· Owl, Fox, Rabbit · agents: BusyBeaver, Ledger
The one rule the beavers never break
No agent knows or does more than the person who asked. Not even by calling another agent.
Most agents answer anyone with everything they can reach, and pass all of it on to the next agent they call. BusyBeaver checks the person behind every request, at every step of the chain.
Owl may deploy.
So BusyBeaver deploys for Owl, once Owl confirms the card.
Rabbit may not see refunds.
So when BusyBeaver calls Ledger for Rabbit, Ledger uses only what Rabbit may see, even though Ledger itself can read refunds.
Rabbit may not deploy.
So no agent does it on Rabbit's word. It asks someone who may, acts only on their yes, and writes both names in the diary.
What just happened in the lodge
01
Owl tells BusyBeaver to deploy a fix. Anyone in the room can direct an agent, and the agent always knows who did.
02
BusyBeaver calls Ledger to look up refunds for Fox. The details go to Fox privately; the room sees nothing extra.
03
When Rabbit asks, Ledger works with Rabbit's permissions, so the ledger stays shut. No agent in the chain ever borrows someone else's rights.
04
Every request, agent call, approval and private note goes into the diary, under the name of the person behind it.
Meet the crew
Every agent works for whoever is asking, and never with more than that person may. Each agent also has its own ceiling.
agent · you talk to it
Takes requests from anyone in the room, always within that person's permissions, and calls other agents when a job needs them.
acts after a yes
agent · a helper
Reads the finance ledger. Called for someone, it uses what that person may see and what its own ceiling allows, whichever is less.
ceiling: finance
person · admin
Cleared for refunds and billing. Can deploy, and approves other people's requests.
sees refundscan deploy
person · analyst
Cleared for refunds, so agents may fetch refund details for Fox. Fox's deploys need an approval.
sees refundsneeds approval
person · contractor
Content only. Can ask any agent anything; refund details never reach Rabbit, whichever agent fetches them.
content onlyneeds approval
The diary
Every question, agent call, approval, private note and admin change goes into a diary, under the name of the person behind it. Each line is chained to the one before, so nobody can quietly edit history.
It records ids and reasons, never the text people typed.
#ops · today
✓ chain verified
For builders
BusyBeaver's core is a permission layer for agent systems. Whichever harness or model you use, every context an agent builds, and every agent it calls, stays within the permissions of the person who asked.
label
A note without an audience can't be stored, so nothing enters a context unlabelled.
admit
Before any model call, each note is checked against the people behind the request, and the agent's ceiling.
inherit
Whatever an agent writes is labelled with the strictest of what it read, so nothing gets laundered down the chain.
log
Admissions, refusals, approvals and agent calls go into a hash-chained log you can verify.
The benchmark
We asked 24 questions in rooms of people with different permissions. The same model leaked restricted information in 18 of them without BusyBeaver, and in 0 with it.
Without BusyBeaver
18/24
questions leaked
With BusyBeaver
0/24
questions leaked
Synthetic company data; methodology available on request.
Admins choose which actions each room allows. A deploy asked for in #finance stops before anyone is asked to approve it.
A note someone may not see looks exactly like a note that does not exist. No "I can't tell you that".
Anthropic, OpenAI, Google or open models through Ollama. The model only gets notes that passed the checks, and it never runs tools itself.
We are working with a few teams that run AI in shared channels. Tell us about yours.