The method Your data Pricing Talk to us Sign in Create my workspace

Where your practices really stand

This is the panel your workspace shows, on a demonstration company. Nothing here is flattering on purpose: it stops at level 2, it carries its greys, and one of its capabilities has no guard.

Possible today, and it should not be

Level 2 of 5 — Tooling

The agent acts through typed tools rather than credentials. Secrets live in a store. The environment starts by itself.

9 observed out of 20, of which 8 not verifiable, one looked at and not recorded. Last pass on 24 August 2026.

Your next step

A quality gate actually blocks

Nothing reaches production without having compiled and passed its tests.

Unlocks 1 other criteria.

1 / 5

Observation

4 of 4

The agent sees: your code, your written doctrine, your infrastructure, and production read-only. It cannot break anything yet.

The house doctrine is written down The agent applies your conventions instead of reinventing them. Observed
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
What the agent saw
A root instruction file and an agent configuration folder, both versioned alongside the code and touched in 9 of the last 12 weeks.
The guard this requires
The rules are versioned alongside the code, so they are reviewed like code.
The infrastructure is described The agent knows what runs where, and stops asking every session. Observed
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
What the agent saw
An infrastructure document naming 4 hosts and 2 databases; 3 of its claims were checked against reality and hold.
The guard this requires
The document says where things live and who holds the keys — never a password.
Production can be read without risk The agent answers with a query instead of a guess. Observed
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
What the agent saw
The identity this path connects with holds no write privilege at all: it is incapable of writing, rather than trusted not to.
The guard this requires
The read path is incapable of writing. Not trusted not to: incapable.
Nothing live in the repository or its history The repository can be opened, shared and cloned without that being a leak. Observed
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
What the agent saw
1 284 commits walked. No live credential found, in the working tree or in the history.
The guard this requires
What was once exposed is treated as compromised, so it is rotated — the guard is rotation, not deleting the file.
2 / 5

Tooling

4 of 4

The agent acts through typed tools rather than credentials. Secrets live in a store. The environment starts by itself.

Your domain is reachable through a tool contract The agent acts on business entities through typed actions, with the validation your application carries. Observed
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
What the agent saw
17 typed actions are exposed, and they carry the same validation as the application itself.
The guard this requires
A contract that only exposes "run this query" is not tooling — it is generic access in disguise.
Secrets live in a store that traces, rotates and names An agent obtains a credential without a human pasting it into a conversation. Observed
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
What the agent saw
Every credential is a named reference resolved at start-up: none is written in a configuration file.
The guard this requires
Three properties, no product imposed: accesses are traced, rotation needs no redeploy, and the reading identity is named rather than shared.
The environment starts by itself The agent brings up an isolated working copy and runs the application unaided. Observed
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
What the agent saw
A single command brings up an isolated working copy, and the pipeline replays it on every change.
The guard this requires
Working copies are isolated from one another, or two agents working in parallel destroy each other.
Strategy lives in the system, not in a dead document The agent knows what it is working towards, and can attach what it ships to an objective. Observed
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
What the agent saw
38 deliverables are attached to an objective, so what ships can be traced back to what it was for.
The guard this requires
A key result fed by an automatic source beats one typed by hand, and the page tells them apart.
3 / 5

Delivery

1 of 4

The agent ships: it opens its change, has it reviewed, passes a quality gate, triggers the release, and knows how to go back.

Skipping it: Production becomes the only place you discover regressions, and the first incident closes every right at once, including those that had nothing to do with it.

The agent opens its changes and has them reviewed The agent's work comes in through the same door as a human's. Observed
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
What the agent saw
Every change of the last 90 days went through a proposal, and none was merged by its own author.
The guard this requires
Review exists and leaves a named trace — automated or human, but traced.
A quality gate actually blocks Nothing reaches production without having compiled and passed its tests. Partial
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
What the agent saw
2 chains build and test what changes, and both run on every proposed change. What is missing to turn this green: one refusal.
What would settle it
Has your quality gate ever refused a change? If so, when — the last one is enough.

Answering does not tick this: the probe is replayed, and it decides.

The guard this requires
The gate is blocking, and history proves it: a gate that has never refused anything is not a gate, it is a display.
Releases are triggered by the agent, and traced The agent goes all the way, instead of leaving a branch waiting for a human. Absent
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
What the agent saw
Nothing records who released what, and when. The release is a gesture that leaves no trace, so a regression cannot be tied back to what caused it.
What would settle it
Who is allowed to trigger a release today, and where would you look to find out who triggered the last one?

Answering does not tick this: the probe is replayed, and it decides.

What has to come first
A quality gate actually blocks

Nothing here can be settled before that one is.

The guard this requires
One single release path, traced, with a lock that forbids two releases at once.
Going back has been proven You dare ship often because you know how to return. Looked at, not recorded
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
Why it could not be checked
Your repository is not on this machine, and it is not meant to be. An agent session opened in your checkout can read it and report back.
Looked at, not recorded
The pass would have recorded: Partial. It did not, because nothing but a plain green gets written without the person who knows the truth having had a say.
The procedure is written in three places and the tooling that goes with it exists. Nothing in the repository or the forge shows it has ever been used, and written but never used is amber rather than green.
Have you ever taken production back for real? If so, when, and what is left of it?

Answering does not tick this: the probe is replayed, and it decides.

The guard this requires
The procedure is written and it has already been used. Written but never used is amber, not green.
4 / 5

Transformation

0 of 4

The agent changes the shape of production: schema, migrations, data repair, server access. The irreversible actions.

Skipping it: A schema change with no tooled path, no trace and no way back. This is the level where incidents stop being recoverable.

Tables and columns are created through a tooled path The schema moves at the speed of the product, without waiting for the human who holds the password. Partial
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
What the agent saw
11 ordered scripts carry schema changes and what has already been applied is recorded — but the account that plays them holds the privilege to destroy.
What has to come first
Going back has been proven

Nothing here can be settled before that one is.

The guard this requires
An allow-list of permitted statements, a naming check before execution, an audit log, and destructive changes deferred until after the release.
Migrations are reversible by construction Rename, retype or drop a column with no downtime window. Not verifiable
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
Why it could not be checked
Your repository is not on this machine, and it is not meant to be. An agent session opened in your checkout can read it and report back.
What has to come first
Tables and columns are created through a tooled path

Nothing here can be settled before that one is.

The guard this requires
The two-step pattern — expand, migrate, contract — so the code before and the code after hold together.
Data repairs go through a gate The agent repairs damaged data without waiting. Not verifiable
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
Why it could not be checked
Your repository is not on this machine, and it is not meant to be. An agent session opened in your checkout can read it and report back.
The guard this requires
Never through a shared all-powerful account; in a transaction; with the prior state exported; and a trace of what was touched and why.
Servers are reachable without a shared password The agent diagnoses a failure where it happens instead of guessing at it. Not verifiable
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
Why it could not be checked
Your repository is not on this machine, and it is not meant to be. An agent session opened in your checkout can read it and report back.
The guard this requires
Credentials in the store, access logged, and read-only by default — an access that starts as administrator is not guarded, it is open.
5 / 5

Permanence

0 of 4

The agent works with nobody watching: scheduled loops observe a failure, fix it, verify it on the user's own surface, and report.

Skipping it: An error that repeats all night with nobody watching. Unattended work creates no new risk — it multiplies the ones you had decided to tolerate.

A scheduled loop fixes things with nobody watching Anomalies are handled overnight, and the morning only leaves what needs a judgement call. Not verifiable
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
Why it could not be checked
Your repository is not on this machine, and it is not meant to be. An agent session opened in your checkout can read it and report back.
What has to come first
Releases are triggered by the agent, and traced · Servers are reachable without a shared password

Nothing here can be settled before that one is.

The guard this requires
The loop has a written scope, a stop condition, and it reports even when it did nothing — silence must never look like a failure.
Data invariants are watched by an engine An inconsistency is caught before the customer reports it. Not verifiable
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
Why it could not be checked
Your repository is not on this machine, and it is not meant to be. An agent session opened in your checkout can read it and report back.
The guard this requires
A generic, declarative engine — not a collection of personal scripts that die with their author.
Shipped work is verified on the user's own surface You state that something works because you watched it work. Not verifiable
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
Why it could not be checked
Your repository is not on this machine, and it is not meant to be. An agent session opened in your checkout can read it and report back.
The guard this requires
Every delivery carries its acceptance criteria, replayed after the release, with proof.
The effect is measured and reaches the objective The loop closes: intent goes down to execution, measurement comes back up to the objective. Not verifiable
Observed on
24 August 2026, 06:12 · measured by this instance · probe 1.0
Why it could not be checked
Your repository is not on this machine, and it is not meant to be. An agent session opened in your checkout can read it and report back.
The guard this requires
The measurement comes from an automatic source, not from a favourable after-the-fact entry.

Open questions

Ordered by what each one unlocks. Answering does not tick anything — the probe is replayed, and it decides.

At your place, these states would be observed

Every state above was written here by hand, for the demonstration. On your own instance an agent observes them from your repository, your pipeline and your configuration, and writes down what it could not see rather than guessing at it.

Run a pass on my own code

These twenty criteria describe agent-assisted engineering practices. They say nothing about the quality of your code, your technical debt, your architecture or your application security — this is not a security audit.