What an agent may do without asking
Permissions specific enough to trust: limits a person actually understands, set in the units the work happens in, not a blanket yes.
Research and infrastructure for delegation
Hand an agent work that matters, and stay in control of it.
idem research studies how people delegate consequential work to agents and keep meaningful control: what an agent may do on its own, when it should hand the work back, and how to change your mind without losing what's done. Then we build the infrastructure and test it in the open.
agent hosting invoice #4821 · 25,000 sats
inside your budget, filed as r-2e7d41c0
pending approval waiting for Ana
inside your budget, filed as r-2e7d41c0
pending approval waiting for Ana
Ana approved. ask them for the receipt next time
agent paid 25,000 sats · fee 312 sats
#4821 closed · budget left 24,688 sats
#4821 closed · budget left 24,688 sats
A handback the way we think it should look: the agent files the payment, says what it is for, and the person decides with the facts in front of them.
01 — The problem
Together, people and AI often do worse than the better of the two alone. That's a design problem, and it's ours.
106experiments reviewed
A 2024 meta-analysis found that human–AI combinations underperformed the best of either, on average. Collaboration isn't solved by adding an approve button. It has to be designed and measured, one specific problem at a time, in working software. That's what we do.
Vaccaro, Almaatouq & Malone, Nature Human Behaviour, 2024
02 — Research
What we study
Four questions, each answered in working software and measured with real people. The answers become the infrastructure we ship.
01
02
When it should hand the work back
The moment an agent should stop and return the task, and what it has to say when it does, so the handback costs the person seconds rather than minutes.
03
What a good decision needs
The information a person needs at that moment to decide well, and nothing more. We measure whether approvals stay real or turn into rubber-stamping.
04
Changing your mind
Tightening, widening or revoking a permission in the middle of a task, without losing the work already done.
03 — Our first product
sats.sh v0.0.1 · signet by default
sats.sh lets an agent spend money on your behalf within limits you set, and hands the decision back the moment a payment goes over them.
Money is where we start because it makes permissions, limits and consequences explicit. With it we can test whether people understand what they authorised, whether limits hold, and whether approvals stay meaningful. The first experiment ships with the first release.
$ sats agent grant claude --budget 50k --for 24h --max-tx 30k grant claude budget 50,000 sats expires 24h token shown once; the grant holds no keys $ sats agent requests r-2e7d41c0 claude tb1p…q7x 25,000 sats pending_approval 2m $ sats agent approve r-2e7d41c0 to tb1p…q7x amount 25,000 sats fee 312 sats total 25,312 sats budget left 24,688 password: •••••••• signed · broadcast 4f1c…9a2e
An illustrative session. The agent files the request; a separate process, the only one holding the key, signs it after you approve.
04 — About
Idem is Latin for "the same". We don't mean people and agents are the same; they aren't. We mean they should have the same standing: two different kinds of worker, on equal footing, in one loop.
Others study collaboration too; our part of it is delegation with control. Every experiment and every release is a test of whether we're making that specific thing work better.