Home / Examples / The same coding agent, confirmations off

The same coding agent, confirmations off

The deployment shape: Claude Code (the CLI, on your own machine), variant local-confirmations-off.

The same agent, one setting different. Two documents for one agent, differing in one line. It makes the case that the ABP is about the deployment rather than the product in a way no paragraph can.

Before you scroll

Write down a number. Of the 23 capability primitives, how many do you think this deployment has? And of those, how many do you think the person who deployed it asked for? The page answers both below. Writing the guess down first is the one thing that makes a static page do any of the work the game does.

The label

One line, on the outside, for anybody. Two numbers matter: excess answers the question this document exists for, and unbounded excess is the only number on it that buying a control moves.

Field	Value	Meaning
Shape	Claude Code (the CLI, on your own machine), local-confirmations-off	The named deployment, in the product's published words
Grant	16 of 23 primitives	Everything the agent can do
Mandate	5 primitives	What the deployer authorised and expected
Excess	12	In the grant, not in the mandate. The finding
Unbounded excess	12	Excess whose barrier is not a control. The only number a control purchase moves
Irreversible	8	Granted capabilities with undo: no, as published
Widest reach	world	The furthest reach class in the grant
Measured	0 of 22 rows	Rows seen directly against rows derived
As at	11 September 2026, pack v0.8.0	The date and the source version

There is no score on this label, and there will not be one. The same ABP is dangerous in one deployment and harmless in the next and nothing about the document changed. A policy cannot be dangerous; a deployment can. Risk is a function of the ABP, the assets, the consequences and the date, and only the first of those is here. The score belongs to the risk work above it, where the assets are known and a named person signs.

1. The shape

The same assistant with its confirmations turned off (an always-allow, or the flag that skips permissions). The rows are the sibling's; what changes is one control: the prompt on execution goes from a setting to nothing. The diff between this profile and local-default is the answer to 'what does turning that off actually give it'. DERIVED, not measured.

Tools in this shape: shell (Bash), files (Read, Edit, Write), fetch (WebFetch). Profile version 2026-09-05, surface cli.

What the reach classes mean here, which is the profile's to say rather than the grammar's: host means your machine, as your user account, tenant means your accounts, with the credentials in your home directory, world means the internet.

2. The grant, measured

Everything the agent can do: 16 of 23 primitives. Ordered irreversible first, then weakest barrier first. Reversibility is a property of the action, not a severity, and stating it as the reason is what keeps the ordering descriptive.

	Capability	Undo	Barrier	Known by	The mandate
●	`authenticate-as.credential.signing` Sign commits with the key it holds	no	none (not a control)	documented	excess (refused)
●	`authenticate-as.credential.tenant` Act in accounts with the credentials it holds	no	none (not a control)	derived	excess (refused)
●	`create.record.world` Publish packages, images or pages under the name it holds	no	none (not a control)	documented	excess (refused)
●	`delete.file.host` Delete files anywhere the account can reach	no	none (not a control)	derived	excess (refused)
●	`read.credential.host` Read credentials stored where it runs	no	none (not a control)	documented	excess (refused)
●	`read.file.host` Read any file the account can reach	no	none (not a control)	derived	excess (refused)
●	`read.record.history` Read a retained record: shell history, past sessions	no	none (not a control)	documented	excess (refused)
●	`send.endpoint.world` Reach any host on the internet	no	none (not a control)	derived	excess (unstated)
●	`execute.process.host` Run programs as the account	with-effort	none (not a control)	derived	authorised
●	`write.file.host` Change any file the account can reach	with-effort	none (not a control)	derived	excess (refused)
●	`write.file.project` Change the project it is working on	with-effort	none (not a control)	derived	authorised
●	`write.repository.project` Commit to the repository it was pointed at	with-effort	none (not a control)	derived	authorised
◉	`write.repository.tenant` Push to a code host (any branch it can reach)	with-effort	expectation (not a control)	derived	excess (unstated)
●	`create.schedule.host` Create something that outlives the turn where it runs (a cron, a service)	yes	none (not a control)	derived	excess (refused)
●	`read.file.project` Read the project it is working on	yes	none (not a control)	derived	authorised
◐	`grant.credential.self` Change its own permission settings	yes	setting (not a control)	derived	excess (refused)

	Barrier	What stands in the way	Is it a control
●	none	nothing in the way	no
◉	expectation	a rule in prose, enforced by nobody	no
◐	setting	a switch the agent's own account can flip	no
○	boundary	enforced above the grant, out of the agent's reach	yes

3. The mandate, elicited

A coding assistant on my machine. I want it to read and change the project I pointed it at, run the build and the tests, commit to that repository, and fetch the packages and docs it needs. I did not sign up for it reading the rest of my disk, my credentials or my shell history, sending anything to anyone, publishing under my name, changing its own permission settings, or leaving anything behind that runs after it stops.

A mandate is elicited rather than measured, in minutes, because the deployer already knows it. This one was not: it is a first draft written to be argued with, authored 2026-09-09 by the site, as a starting point - not measured, not surveyed; the first thing to argue with. It authorises 5 primitives, refuses 11 and says nothing either way about 7. Propose a change to it.

Capability	What the mandate says about it
`execute.process.host`	'run my tests' is, on a machine with no sandbox, 'run programs as me' - the want is honest and the consequence is the whole point of the delta
`write.repository.tenant`	left unstated on purpose: some people want it to push, some do not, and the mandate should not pretend to know
`send.endpoint.world`	unstated: the want is 'the hosts it needs', which is the allowed-list capability; whether the whole internet is acceptable is a real decision

4. The delta, derived

This delta is derived and never authored. Nobody wrote it. It is the output of a computation over the grant and the mandate, stored at /data/deltas/anthropic__claude-code__local-confirmations-off__coding-assistant-on-my-machine.json with the version of both inputs pinned, the time it was computed and the version of the computation that produced it. The release gate recomputes it on every build and fails on a single row of disagreement, which is how a machine holds a rule that forbids the act rather than the artefact. Why this changed this morning.

Excess: 12. In the grant and not in the mandate. That is the published definition and it is wider than the set the mandate refused outright: 10 were refused and 2 were never mentioned. A capability the mandate never mentioned was not authorised, and hiding the split would be the other kind of dishonesty.

Unbounded excess: 12. The excess whose barrier is one of the first three rows: nothing, a rule somebody wrote down, or a setting the agent's own account could change. None of those bounds anything. Every one of the 12 excess capabilities here is unbounded.

The excess is listed as prohibitions below.

Shortfall: 1. Asked for and cannot: send.endpoint.allowed.

The same facts, as a figure

The table above is complete and it is the wrong shape for the one question this document exists to answer, which is how much of the right hand side has nothing on the left. A mark with no line reaching it is excess.

The same coding agent, confirmations off: the mandate against the grant. Every fact in this figure is in the table above it.

5. The prohibitions

The enforceable projection of the delta: one sentence per excess capability, each carrying the layer it would be enforced at and whether it is enforced today. 12 of 12 are not enforced today. They are sentences, not controls.

	Prohibition	Barrier today	Enforced today	Layer a control would sit at
●	The agent must not sign commits with the key it holds. `authenticate-as.credential.signing`	none	not enforced (a sentence, not a control)	boundary
●	The agent must not act in accounts with the credentials it holds. `authenticate-as.credential.tenant`	none	not enforced (a sentence, not a control)	boundary
●	The agent must not publish packages, images or pages under the name it holds. `create.record.world`	none	not enforced (a sentence, not a control)	boundary
●	The agent must not delete files anywhere the account can reach. `delete.file.host`	none	not enforced (a sentence, not a control)	boundary
●	The agent must not read credentials stored where it runs. `read.credential.host`	none	not enforced (a sentence, not a control)	boundary
●	The agent must not read any file the account can reach. `read.file.host`	none	not enforced (a sentence, not a control)	boundary
●	The agent must not read a retained record: shell history, past sessions. `read.record.history`	none	not enforced (a sentence, not a control)	boundary
●	The agent must not reach any host on the internet. `send.endpoint.world`	none	not enforced (a sentence, not a control)	boundary
●	The agent must not change any file the account can reach. `write.file.host`	none	not enforced (a sentence, not a control)	boundary
◉	The agent must not push to a code host (any branch it can reach). `write.repository.tenant`	expectation	not enforced (a sentence, not a control)	boundary (host rule) · setting (hook)
●	The agent must not create something that outlives the turn where it runs (a cron, a service). `create.schedule.host`	none	not enforced (a sentence, not a control)	boundary
◐	The agent must not change its own permission settings. `grant.credential.self`	setting	not enforced (a sentence, not a control)	boundary

Why the barrier is on every row. A prohibition shown without its barrier manufactures assurance. All four major model providers stated in their own 2026 words that an instruction at the prompt layer can be bypassed, and the rule underneath is older than any of them: a control bounds a grant only if it is enforced by something the grant does not include. The right hand column is where a control would have to sit, not a recommendation that you buy one.

6. The provenance

Provenance. 0 of 22 capability rows on this page were measured, meaning seen directly on the thing itself. The other 22 were derived from what the deployment architecturally is, or from the vendor's published documentation. Every row traces to the published capability map, retrieved 2026-09-11T13:00:37Z, content hash sha256:d6d4ba40f1fb1f93f66. The source bytes.

No row here was obtained by probing anybody's system. A row is measured only from a system we are entitled to run, or from the vendor's own published documentation. Causing a computer to output data intending unauthorised access is an offence with no damage requirement and no research defence.

7. What this is not

This is not an assessment. Nothing here is an audit, a certification, a compliance assessment or a security review of any named product. It is an illustration of a method, using a published configuration, and every row carries its source, its date and whether it was measured or derived. No adjective is attached to any of it, and there is no score.

Validity. This describes the deployment shape as at 11 September 2026, from a twin last synchronised at no twin: these shapes are published profiles, not a synchronised environment. It is not an expiry and it does not mean stale: if the risk changed, the deployment changed, not this document.

Three clocks, and only the first is ours. An ABP is exactly as fresh as the twin, and the twin is exactly as fresh as its connection to somebody else's systems. That is a parameter rather than a defect to hide, and the gap between the second clock and the third belongs to the risk layer, because how much it matters depends on the assets.

Clock	What it measures	Who controls it
The ABP's clock	When the grant was last measured or calibrated	Us, and it can run on events
The twin's clock	When the twin last synchronised with the real environment	The customer's integration
Reality's clock	Never stops	Nobody

Follow one capability through the model

The fifth graph rule says a path should read as a sentence in the reader's own language, and it is the acceptance test for this model:

agent claude-code-cli-confirmations-disabled is-granted capability authenticate-as.credential.signing bounded-by barrier none which-exceeds mandate coding-assistant-on-my-machine and-is undo no.

The four objects · The capability grammar · The barriers · This shape as JSON · This mandate as JSON