Is autonomy earned or assumed?
For every agent acting in your name, which human owns the outcome, what did it cost, what evidence would justify letting it act alone, and are you collecting that evidence, or hoping?
Why this question
Being able to do the work should not automatically mean being allowed to. That sentence is the entire architecture. Agents can now take a ticket, do the work and hand it back. What they cannot supply, and what no protocol and no vendor supplies, is the answer to the questions anyone carrying risk asks first: who is this, on whose authority, what did it do, what did it cost, and who approved it.
Autonomy is a grade, held per agent and per jurisdiction, earned against criteria published before anyone started measuring, and lost automatically. The business owns four things that make that possible and assembles everything else: what it knows (the graph), what acts in its name (the register), the bar each agent has to clear (the gates), and the tests built from its own files (the harness). This question walks through the four, the plan that puts them in place in ninety days, and the mandate that owns them.
It is the longest of the three questions, ten concepts, because it is the one you can act on this quarter.
- The tell. Software went first. Nobody will have one agent. Why the orchestration layer is the position.
- Settled and unsettled. Three protocols done, five questions unanswered. The gap.
- The boundary. Four owned things, and the line where vendors stop. The architecture in one picture.
- The graph. What the business knows, arranged so agents can use it. The first owned thing.
- The register. Ten fields, fifty lines, one screen a board can read. What acts in your name.
- The gates. Observe, recommend, prepare, act. Demotion is automatic. How autonomy is earned.
- The harness. Your files, your judgement, held back from the vendor. What every grade is earned against.
- Least intelligence. The cheapest model that passes, per step, and bytes that left your control. The economics and the privacy number.
- Ninety days. Six artefacts, three people, a portfolio run like products. The plan.
- The mandate. Whoever holds it owns five things. The owner.
- The Autonomy Map. Autonomy granted against evidence held. Piloting, Assuming, Earning, Earned. Up before right. A firm-level picture; the Ladder is the agent-level one.
- The Boundary. Rented, assembled, owned, and the vendor line that moved inward this year.
- The Ladder. Four grades and one arrow back.
- The Grade Board. Agents down, jurisdictions across, the grade in each cell.
The map is on this page. Two questions place you. Most firms are piloting: copilots everywhere, nothing in production with an owner. The dangerous quadrant is assuming: agents acting under people's own logins with no register and no test set. The path from either to earned goes up first, building evidence, before it goes right. The ninety-day plan is that path drawn on the map.
This is a software story and it doesn't transfer (The tell; conceded in part). The standards will fill the gaps within a year (Settled and unsettled; conceded in part). Earned autonomy is in an AWS report (The boundary; conceded in part, and welcomed). This is an ontology project by another name (The graph; held). This is just a CMDB for agents (The register; held). Staged autonomy slows everything down (The gates; held). Vendors ship evaluation frameworks now (The harness; held). Cheaper models mean worse outcomes (Least intelligence; conceded in part). Ninety days is a consultant's number (Ninety days; held). Chief AI Officer is a fad title (The mandate; conceded in part). Each is answered in full inside the concept it attacks.