Agent Authority Authority to act
Every agent action traces back to a person, within limits that only narrow.
Coming soon Agent Authority is coming soon. Nothing on this page is available yet.
A person grants an agent authority. Every hand-off passes on less, never more, and every action is checked before it reaches your systems.
On the record. Every grant, action and refusal down the chain lands on a record no agent can rewrite.
The spread
One approval. Everything it starts. Your systems.
A person approves one agent. That agent starts as many sub-agents as the work needs, and each one inherits the reach it was given.
Nobody approved the rest
A person looked at one agent. Everything it starts inherits that reach, and no one is asked a second time.
They outlive the task
A sub-agent that finishes its work keeps what it was handed until somebody goes looking for it.
The record says one name
Every call arrives under the same credential, so the log tells you a person acted and nothing more.
More rules do not change the spread. A limit carried by everything the grant starts does.
What it returns
The answer is accurate. The account is still missing.
The agent writes its own account, and nothing outside it verifies the accuracy.
What it leaves you unable to say
- What work actually happened, and from what source
- Whether data left the building
- Whether anything was refused, and on whose authority
An account you cannot check is not evidence. The record has to be written outside the system that acted, where the agent cannot reach it.
The answer Coming soon
Eight problems. One place to stop them all.
Every one of them
- Overreach
- Hijacked agents
- Runaway agents
- Stolen credentials
- No way back
- No trusted record
- Unknown agents
- Endless authority
Each one is an action taken with a person's authority, unlimited and unrecorded.
Meets the same three things
01A person's grant
Authority comes from a named person and only narrows as work passes down.
02A limit no agent can raise
Each agent can do only so much, and no agent can give itself more.
03A record out of reach
Written once and held where the thing being recorded cannot edit it.
We do not solve eight problems eight ways. Every action meets one gate, and the gate answers all eight.
The outcome
Identity says who it is. The record says what it did.
Every action is tied to the person who allowed it and recorded as it happens. None of it is accounted for by the agent.
| Who | Action | Outcome |
|---|---|---|
| sub-agent 1 | read CRM | Allowed |
| sub-agent 4 | update ticket | Allowed |
| sub-agent 2 | send email | Allowed |
| sub-agent 4 | export contacts | Allowed |
| caller, unverified | request customer records no grant, flagged | Refused |
Without it, the row names whatever identity the call used, a shared agent or the person's own token, and the refusal isn't there at all.
01Actions cannot hide
Nothing happens outside a person's grant, so every action traces to the person who granted it.
02Your own AI gets protected too
A manipulated agent can be talked into attempting something. It cannot be talked into authority it was never granted, and everything it tries is on the record.
03You stop taking its word
The reason for an action is recorded as it happens, not narrated by the agent afterwards.
Four questions
The same four questions, answered.
| Problem | The question | The answer |
|---|---|---|
| 01Runaway | How many changes can one agent, and the sub-agents it starts, make in our systems? | No more than a person granted, and it stops there. |
| 02No way back | When we revoke a person’s access, do the agents acting for them lose it at the same moment? | Yes. One revocation stops every agent and sub-agent acting under that person’s grant. |
| 03No record | If an agent is compromised, what did it do, and under whose authority? | Every grant, action and refusal is on a record the agent can’t reach, so you can trace each step back to a person. |
| 04Overreach | When did an agent first reach sensitive data, and whose grant let it? | A sub-agent holds only what it was handed. The first reach of sensitive data is flagged to the person whose grant allowed it. |
One design answers all four.
What you can do
Know what your agents can do, limit it, and prove it.
01Know
See every agent and whose authority it acts on.
02Limit
Limit how much each agent and its sub-agents can do.
Stop every agent acting for a person, at once.
03Prove
Prove what happened without taking the agent's word.
Your own team, your auditors and your customers.
Design partners
Agent Authority is coming soon.
We are forming a small group of design partners who want these questions answered, and whose environments shape the final product. The way to start today is an Agent Access Review: two weeks to see which agents are running, what each can reach and whose authority they act on.