Mission Control for Long-Running AI

Replace vague loading with a real task workbench.

`agent-progress-ui` is an open-source React UI system for deep research, code agents, and long-running AI tasks. Show liveness, stages, and proof instead of trapping users behind a spinner.

Liveness

Always prove the task is still alive.

Stages

Stable checkpoints instead of fake progress bars.

Evidence

Sources, tools, drafts, and reviewable artifacts.

Live Scenario

Agent Workbench

2-10 minutes. This is no longer a loading state.

Synthesizing01:34
Running checks

Build and smoke checks are underway.

CancelBackgroundReview

Stage Rail

Plan executioncompleted

Lock scope, constraints, and success criteria.

Inspect workspacecompleted

Read files and infer the safe edit boundary.

Apply changescompleted

Mutate the project with checkpoints.

Run checksactive

Execute tests and gather proof.

Package outputpending

Collect diffs, screenshots, and review notes.

Artifacts

diff

Patch preview

A reviewable diff summary with ownership notes.

Evidence Feed

Running checks01:34

Build and smoke checks are underway.

Implementation checkpoint saved01:32

The code changes are stable enough for verification.

Task moved to background01:10

The user can leave without losing progress.

Artifact created01:04

A patch preview is available before verification.

Applying changes00:49

The agent is editing and staging artifacts for review.

Inspection complete00:48

Safe edit boundaries are now visible.

Long-running AI work needs observability, not decoration.

Once a task crosses the 10-second mark, users stop asking “is this pretty?” and start asking “is this alive?”, “what is it doing?”, and “can I safely leave?”. This project turns those questions into reusable UI primitives.

Anti-pattern

A single spinner that runs for 30 seconds with no evidence.

Anti-pattern

A fake 64% progress bar for work that depends on external sources or tools.

Anti-pattern

One visual treatment for executing, waiting on rate limits, and waiting on the user.

A system, not a spinner.

The package is organized around the three surfaces that matter in long-running AI experiences: keep the task alive, show the stage, and emit proof as the system learns or changes the world.

Liveness Layer

Every task proves the system is still alive without turning the interface into a spinner prison.

Stage Layer

Progress is communicated as a stable sequence of stages instead of fake percentages.

Evidence Layer

Users see sources, tools, drafts, and artifacts as they appear, which is how trust is actually built.

Stable states are more trustworthy than fake percentages.

`agent-progress-ui` ships with a fixed state machine, reducer, and scenario catalog so teams can design around real execution phases instead of ambiguous “thinking” copy.

queuedunderstandingplanningexecutingsynthesizingwaiting_externalwaiting_userretryingbackgroundedready_for_reviewcompletedfailedcancelled

Event schema sample

{
  "id": "agent-15",
  "taskId": "task_agent_01",
  "type": "step_completed",
  "state": "ready_for_review",
  "timestamp": "2026-04-29T09:02:02.000Z",
  "title": "Ready for review",
  "message": "Diffs, artifacts, and next steps are bundled for inspection.",
  "stepId": "agent-verify",
  "stepLabel": "Run checks",
  "elapsedMs": 122000,
  "artifact": {
    "type": "bundle",
    "id": "agent-review",
    "label": "Review bundle",
    "description": "Diff, screenshots, logs, and verification output."
  }
}

Use the same system across chat, research, and agents.

Open playground