Progressive Autonomy
The ladder you design is per scope; every shipping harness's dial is per session — so the rung has to live in the permission rules, and the mode selector only decides who gets asked.
Claude Code ships six permission modes as of October 2026 and Codex CLI crosses three sandbox modes with a separate approval policy, and not one of those settings can express "act unsupervised when labelling tickets, ask every time when issuing refunds" at the same moment — because they are all session-global. Ship the autonomy ladder as a mode selector and you have built a product where the user must pick one trust level for two different risk profiles, which means they will either over-constrain the safe task or over-trust the dangerous one. The rung belongs in the allow/deny rules and the sandbox boundary, promotion belongs to a visible track record within one scope, and demotion has to cost less than promotion or nobody will rationally climb.
Step 1The ladder is a sequence of human roles, not agent power levels.
Describe each rung by what the human does, because that is what the user actually experiences and what they are consenting to when they move:
- Operator — human drives; the agent only executes explicit instructions.
- Collaborator — the agent proposes and acts on small steps; the human works alongside.
- Consultant — the agent recommends; the human decides and executes.
- Approver — the agent does the work; the human approves before effects land.
- Observer — the agent acts; the human monitors and audits after the fact.
Higher is not better. A rung is appropriate or not for a given scope and a given track record, and the product's job is to make the current rung legible and the next rung's preconditions explicit. The vocabulary comes out of the 2025 levels-of-autonomy literature and it has held up as a way to talk to users — "you will approve each refund before it goes out" is a sentence a customer-support manager can evaluate, where "autonomy level 3" is not.
What has not held up is the assumption that this ladder maps onto a control in the software. It does not, and the next step is the evidence.
Step 2What the harnesses actually ship is a session-global dial.
Look at the two most widely deployed agent harnesses as of October 2026 and the pattern is identical: a single setting that applies to the whole session, orthogonal to anything about the task at hand.
Shipping autonomy controls, October 2026. Scope of each: THE SESSION. Claude Code — six permission modes default reads auto-approved; edits and commands asked acceptEdits reads + file edits auto-approved; commands asked plan read-only; produces a plan for approval auto a separate classifier model reviews each action dontAsk auto-denies anything not pre-approved (CI shape) bypassPermissions no prompts; documented for containers and VMs Codex CLI — two orthogonal dials sandbox read-only | workspace-write (default) | danger-full-access approvals when the agent asks, set independently of the sandbox note network off by default under workspace-write Neither can say: Observer for "label ticket", Operator for "issue refund".
Two details in that table are doing more work than they look. First, Codex's separation of sandbox from approvals is the right factoring and worth copying: the approval policy decides who is asked, the sandbox decides what is possible, and setting approvals to never does not by itself grant unrestricted execution. Second, Claude Code's auto mode is a different kind of thing from the other five — it is described as delegating the per-action decision to a classifier rather than to a static rule, which is the subject of step 4.
Both products also ship organisational overrides, and those are where scope actually appears: Codex's managed configuration can block --sandbox danger-full-access outright, and Claude Code's cloud sessions ignore a defaultMode of bypassPermissions or dontAsk coming from checked-in settings files, so a repository cannot start a cloud session in bypass mode. That is a ceiling, not a ladder — but it is the correct shape for a ceiling, and it tells you where the enforcement layer lives. Lifecycle hooks and harness config covers that surface in detail.
Step 3Bind the rung to a (capability, scope) pair and enforce it below the harness.
An agent can deserve Observer-level autonomy for "label incoming tickets" and Operator-level for "issue refunds" at the same instant. A single global dial forces the user to choose one, and whichever they choose is wrong for one of the two tasks. So the rung cannot be the dial. It has to be a function of the action the agent is about to take.
# The rung is a function of (capability, scope) — never a global flag. level = policy.autonomy_for( capability="refund", scope={"max_amount": 50, "region": "EU"}, ) if level >= APPROVER and request.amount <= 50: await act_then_review(request) else: await recommend_only(request) # Three layers, and only the first two are yours to design: # 1. sandbox / credential scope -> what is POSSIBLE at all # 2. allow-deny policy per call -> what runs without asking # 3. harness mode (session-wide) -> WHO gets asked when 2 says ask # A ladder implemented at layer 3 is a ladder with one rung.
In practice layers 1 and 2 are where the rung actually lives, and both are things you already have primitives for. The credential the agent holds decides the ceiling — a token scoped to read-only on the billing API cannot issue a refund at any permission mode, which is why blast radius is set by the credential and not by the prompt. The per-call allow/deny rules decide what proceeds silently, and they can name the capability and its arguments, which is the thing the session mode cannot. Policy as code for agents is the long form; task scope is the vocabulary for writing the scope half of the pair.
The design consequence for the interface is small and specific: the mode selector stays, but it stops being the autonomy control and becomes the routing control. "Who gets asked, and how" — this user inline, a reviewer in a queue, a second model, nobody — is a genuinely session-level question, and it is the right question for a session-level widget to answer. The whether stays in policy.
Step 4Promotion is gated on a visible track record — and a classifier rung is a rung whose error rate you now own.
Do not let users jump to high autonomy on day one, and do not promote silently. Gate each step up on demonstrated reliability within that scope: a minimum number of executions, an acceptable correction rate, zero reverted high-consequence actions. Make the offer explicit and evidence-backed — "Over 80 refunds at Approver level you adjusted 2 and reverted 0. Move EU refunds under 50 to Observer?" — so the user is consenting to a calibrated change rather than flipping a switch on faith. The track record from designing for trust is the currency that buys a rung here.
Two shipped products in October 2026 took a different route to the same place, and both are worth understanding because they will become the default pattern. Notion's agent release of 19 October introduced a "Review for me" mode in which a separate model checks the agent's actions and asks the user to approve only the ones it flags as potentially unsafe. Microsoft announced controls on the same day that reduce what an agent may do as its assessed risk rises. Claude Code's auto mode is the same idea inside a coding harness.
This is real progress — a classifier can express per-action judgement that a session mode cannot, which is precisely the gap step 3 identified. But be clear about what you have bought. A classifier rung converts an explicit policy decision into a prediction with a false-negative rate, and that rate is now your product's safety margin. Which means it needs the treatment any other detector gets: a measured confusion matrix on your own action distribution, a threshold chosen against the asymmetry between a wrongly-approved refund and a wrongly-blocked ticket label, and monitoring for drift as your action mix changes. Risk-scored permissions covers the operational side, and decision models covers why a confidence number you did not calibrate is not a threshold you can set.
Offer promotion at the moment the evidence is freshest — right after a clean streak in that scope — not buried in a settings page the user visits once. And state the evidence in the offer. A promotion prompt with no numbers in it is asking for faith, and the user who grants it on faith is the user who revokes it on the first surprise.
Step 5Demotion must be automatic, reversible, and cheaper than promotion.
A ladder with no down-rung is an irreversible grant in slow motion. A material incident, an error-rate regression, a scope change, or a context shift must be able to drop the agent a rung automatically — continuous verification rather than a one-time graduation. Frame it as a normal safety response rather than a failure verdict: expected, logged, explained, and re-climbable on a fresh track record. An agent the user knows can be safely demoted is one they can rationally promote.
The asymmetry is the load-bearing part, and the shipping harnesses already demonstrate it. Claude Code will not let a session switch into bypassPermissions part-way through unless it was enabled at launch, and dontAsk is deliberately kept out of the keyboard cycle so it has to be set with an explicit flag. Dropping to a stricter mode, by contrast, is immediate. Copy that rule: descending is one action with no confirmation, ascending requires evidence and a deliberate gesture. Where the mechanism needs to persist, write the demotion into the policy layer rather than the session, or it evaporates the next time the agent starts.
Two things make a demotion survivable rather than infuriating. It needs a stated re-entry condition at the moment it happens — "back to Approver after 20 clean refunds" — or the user experiences it as a punishment with no appeal. And the work in flight needs somewhere to go: an action the agent was about to take unsupervised becomes an action waiting for approval, not an error. Undo and reversibility covers the effects that already landed, and permission grants and revocation covers the grant surface the demotion is actually editing.
Step 6Make the current rung and its boundary impossible to misread, then do the expected-value arithmetic before you climb.
The largest autonomy failure in the field is ambiguity about what the agent is currently allowed to do on its own. The user must be able to answer, without digging: at what rung is this agent, for this scope, and what is the exact boundary at which it will stop and ask. Render the rung and its scope at the point of action rather than only in settings — autonomy the user has lost track of produces both over-trust, where they assume it will ask, and under-trust, where they assume it will not. This is also the one place where the session mode genuinely belongs on screen, because when the policy says "ask", the mode is what determines whether anyone will see the question. Approval and confirmation UX has the mechanics of the prompt itself; plan review and editing covers the plan-mode rung specifically.
Then the decision to climb. It is an expected-value comparison and it is usually made without one, so write it down:
Should "EU refunds under 50" move from Approver to Observer? volume 400 refunds / month approval time 45 seconds each time saved 400 x 45s = 5.0 hours / month observed error rate at Approver 2 adjusted in 80 = 2.5% errors that reach the customer 400 x 2.5% = 10 / month cost of one escaped bad refund 50 refund + 25 handling + goodwill 10 x 75 = 750 / month of error cost against 5 hours saved. Climb only if 5 hours is worth more than 750 — and note the error rate was measured WITH a human catching them.
That last line is the trap. A correction rate observed at Approver level is the rate at which a reviewer caught something, which tells you how often the agent was wrong but not how often it would have been wrong unobserved. Treat the measured rate as a floor, size the first Observer window small, and keep the reviewer sampling a fraction of actions after promotion so the number keeps being measured. Decision receipts and audit is what makes the retrospective possible at all.
If the cost of a rare miss at the higher rung exceeds the cumulative time saved across all the successes, stay one rung lower — autonomy that is net-negative on expected value is not a maturity milestone, it is a deferred incident. Three things to build, in order: scope the credential so the ceiling is enforced below the harness; put the rung in per-call allow/deny policy keyed on (capability, scope) and leave the session mode to decide who gets asked; and make demotion automatic, immediate and carrying its own re-entry condition. If you adopt a classifier rung — Notion's reviewer model, Claude Code's auto, a risk score — measure its false-negative rate on your own actions before you let it replace a human approval, because from then on that rate is your autonomy policy. See the lethal trifecta for the combination of capabilities that should never sit above Approver regardless of track record.