Progressive
Autonomy is not a one-time setting. It is earned in steps, and every step can be reversed when the evidence changes.
Founding manifesto
Progressive Autonomy Architecture
PAA is the architecture for earned, scoped, reversible autonomy. Every move toward more autonomy is earned by evaluation, and every move can be reversed by evaluation.
The claim
Teams are already doing progressive automation. They escalate to a human when a model is unsure. They let an agent read freely but gate every write. They start with human approval on every step and remove that human only when it stops catching anything. The pattern is already there, but the implementation standard is not.
Progressive Web Apps did not invent offline caching or responsive layout. They named the target, defined what qualifies, and gave people a shared standard to build toward. PAA is the same move for automation: a vendor-neutral specification that says what a progressively autonomous task must declare, record, and prove.
The primitive is not the agent and not the level. It is the evaluator.
What PAA is
Autonomy
The property being dialed, not a final destination. Many tasks should stay partially governed forever.
Architecture
Observability and gating are designed into the task boundary so measurement, substitution, and control stay tied together.
The spectrum
No automation is the starting condition, not a level. The spectrum begins at the first automated action and runs toward full autonomy as a gradient, not a staircase. Assisted, HITL, HOTL, and Autonomous are reference points along that axis.
- Assisted Human-led with automation support.
- HITL Human approves each action.
- HOTL Human samples or monitors.
- Autonomous Automation acts within guardrails.
The middle matters most. Assisted and Autonomous are the easy ends. The discipline lives in HITL and HOTL, where the task either earns the move from review-everything to monitor-and-intervene or gets demoted back when the evidence weakens.
The mechanism
Promotion and demotion are not exceptions to the system. They are the system, and demotion matters as much as promotion.
| Move | What it means |
|---|---|
| Promotion | A task moves toward more autonomy only after evaluation clears the bar across a meaningful window. |
| Demotion | A task moves back toward oversight when the evidence degrades. Reversal is a first-class part of the system. |
| Instrumentation | A task is not eligible until its inputs and outputs are visible at a typed boundary. |
| Swappability | The gate and instrumentation stay fixed while the implementation behind them can be swapped cheaply. |
| Experimentation | The default is to change one variable at a time so the available data is concentrated on the question that matters. |
A task is not eligible for the spectrum until it is instrumented at its boundary. If the inputs and outputs are not observable, the task cannot be measured, gated, or responsibly automated.
Implementations behind a gate must be swappable. If the substrate is welded in place, demotion becomes a rewrite instead of an operation and the loop stops being real.
What PAA is not
PAA is not the one-shot autonomous app: the impressive demo that nobody can account for. It is the discipline that makes autonomy accountable. Confidence is not an eval either. Self-reported uncertainty is a cheap signal, but the gate must be external.
What PAA provides, and what you own
PAA governs the mechanism, not the threshold. It provides the spectrum, the regions, the promotion and demotion loop, the instrumentation criterion, and the roadmap for moving a task safely toward autonomy. You provide the bar.
That division keeps the framework portable. The standard stays fixed even as the risk profile, domain knowledge, and regulatory constraints change from one task to the next.
Prior art and grounding
PAA does not claim that progressive autonomy is new. It takes the lineage that already exists in driving automation, in safety-critical operations, and in the recent wave of agentic systems, then adds a vendor-neutral implementation layer beneath that convergence.
The novelty is making the evaluator first-class: a concrete object that governs movement along the ladder, carries its own evidence, and can itself be promoted or demoted when the evidence changes.
Why now
The distance between experimenting with agents and trusting them in production is still the central unsolved problem in applied AI. PAA exists to make that distance measurable, governable, and reversible.
The pattern is already out there. What it needed was an architecture.