The Autonomy Gate is a specialist judge. Describe any workflow and it issues a binding verdict — automate, supervise, stabilize, or refuse — then hands you the execution artifact built to walk into a meeting and be acted on.
Capability is cheap now. Authority is not. The question that stalls every automation project isn't whether the model is good enough — it's who is accountable when it acts, whether the action can be undone, and whether it should have been a machine's call at all.
The Gate answers that question and refuses, structurally, when the honest answer is no. Other tools route. This one blocks.
Plain language. The Gate normalizes it into an intake snapshot and identifies the terminal action — the last thing that actually executes.
Reversibility, observability, exception rate, cost of failure — then an adversarial check and five hard gates that cannot be overridden.
A decision packet and the matching execution artifact — setup brief, control plan, governance memo — ending in a build handoff pack.
Every artifact ends in an Operator Disposition. The Gate may recommend — it may never pre-approve. A human signs for the build.
Most systems only know how to say yes. The proof that this one judges rather than sells is that it will tell you no — by name, with the gate cited, and no path to override it.
AI executes without a human approval checkpoint. Bounded, observed, reversible, owned.
AI prepares the work; a named human approves before anything executes. The checkpoint is the control.
The process is too unstable for any autonomy level. Stabilize and produce evidence before reassessment.
Judgment or risk that cannot be delegated — a structural block. When the terminal action is an irreversible external commitment or regulated personal data, the verdict is HUMAN_ONLY and the artifact is a governance memo, not a build.
Two live workflows. The first earns AUTONOMOUS — a bounded KPI narrative that's reversible, observed, and owned. The second triggers GATE-2 and returns a governance memo, not a build. The Gate says no precisely — names the mechanism, preserves the human boundary, and still gives the operator a usable way forward.
Not demos — committed runs. A clean autonomous pass, a supervised pass with a blocking checkpoint, and a hard-gate refusal that cites GATE-2 by name. Open any one to read the full artifact.
Every workflow produces exactly one execution artifact, named by its verdict. Each is built to walk into a meeting and be acted on without explanation. Select one to open the full document.
A bounded Claude Project that produces a reviewed internal deliverable from operator-supplied exports.
A scheduled, unattended Cowork project — folder structure, run cadence, and step sequence for the builder.
A code-first pipeline — trigger, inputs, execution sequence, error handling, and recommended stack.