Agent-First UX Patterns — 4 Principles Human supervision of autonomous systems
01
Graduated Trust
Pause on consequential actions. Give the human a moment to approve, edit, or reject before the agent commits.
approve · edit · reject-with-feedback
02
Legible Autonomy
Show what the agent is doing in plain language, step by step. Raw logs and telemetry available behind a toggle.
step timeline · plain language · log toggle
03
Outcome Over Process
Surface what the agent produced, not how it got there. Per-conversation workspace with agent-declared deliverables.
artifact cards · workspace · deliverable declaration
04
Accountable Handback
End every session with a summary card: what I did, what I made, where it is. The agent is accountable for its output.
summary card · completion signal · handback
Demonstrated in three test scenarios: HITL approval · error recovery · reject-and-steer · Porting to icuboid Studio Session 4

The question

Designing for human supervision of AI is not primarily a safety problem. It is a design problem. Autonomous agents need to communicate what they're doing, pause at consequential moments, and hand back clearly when they're done. These are UX decisions, not engineering ones — and they're almost never designed with the same rigour applied to any other interaction.

A1OS was built to make these patterns concrete. An abstract principle like "graduated trust" is easy to agree with in theory and impossible to evaluate until it's built into an actual interface that actually runs. You can't design supervision patterns in a whiteboard session. You have to see what they look like when an agent is actually doing something and a human is actually watching.

What it was

A1OS was an OS-metaphor for agent orchestration. Local development prototype running on Gemini API at port 3000. A glassmorphic cockpit dashboard showing agent processes, an episodic log with token and cost telemetry, a sandboxed file and shell execution environment, and a Human-in-the-Loop (HITL) interceptor that paused on consequential actions.

The OS metaphor was deliberate: agents as processes, resources as managed allocations, supervision as a first-class operating system concept rather than a bolt-on safety layer.

Three test scenarios

The prototype was built around three demo scenarios — each designed to exercise a different supervision pattern:

Scenario 1
Script execution with approval gate
Agent writes a Python script and attempts to execute it. The HITL interceptor pauses before execution, surfaces an approval card with the proposed command, and waits. The human approves, edits the command, or rejects with feedback. Exercises: Graduated Trust.
Scenario 2
Autonomous error recovery
Agent runs a script, encounters an error (missing package), reads the stack trace, installs the missing dependency, and retries — without human intervention. The step timeline shows each action in plain language. The completion card shows what was produced. Exercises: Legible Autonomy + Outcome Over Process.
Scenario 3
Action rejection with feedback
Human rejects the agent's proposed action and provides corrective feedback. The agent incorporates the feedback and proposes an alternative. The cycle continues until the human approves. Exercises: Graduated Trust + the feedback loop that makes supervision genuinely useful rather than a binary stop/go.

The four patterns

The prototype demonstrated four named patterns for human-AI supervision. These are now documented as a framework: Agent-First UX Patterns →

Graduated Trust
Autonomy is a spectrum governed by action consequence. High-risk actions pause for human approval. Routine actions proceed without interruption.
Legible Autonomy
Plain-language step timeline communicates what the agent is doing. Maximum detail is not the same as legible transparency.
Outcome Over Process
The interface's primary metaphor is deliverables, not steps. Artifact cards surface what was produced, not a transcript of how.
Accountable Handback
A designed moment that explicitly returns control to the human. Summary: what I did, what I made, what I'm uncertain about.

Why it was frozen

A1OS was frozen as a prototype exhibit in June 2026. The decision: it was too technical as a standalone product, its outputs were buried in a shared sandbox, and it overlapped significantly with the agent environment I was building in icuboid Studio.

Freezing it was the right call. The prototype served its purpose — it made abstract supervision patterns observable and evaluable. Now those patterns port into a context where they'll be used, not just demonstrated.

The A1OS prototype still runs. It exists as the "exhibit" for the patterns — a record of what was built to arrive at the framework. The three demo scenarios can be run to show, not just tell, what Graduated Trust and Accountable Handback look and feel like in practice.

What connects to CareLogic

The parallel between CareLogic and A1OS became clear in retrospect. Both are about earned trust at infrastructure scale. CareLogic's protocol lifecycle makes clinical guidance trustworthy by making its governance visible. A1OS's supervision patterns make autonomous agents trustworthy by making their actions visible and interceptable.

The question in both cases: what does the system need to expose so that the human who depends on it can act with appropriate confidence?

Tech stack

Local dev · port 3000 Gemini API Sandboxed file + shell execution HITL interceptor (risk heuristic) Glassmorphic cockpit dashboard

Where the patterns live now

icuboid Studio — Session 4 → Full framework documentation →
← Previous
Agentability.io
Ask Rohan
Powered by reflective memory · Phase 1