Make agent work reviewable.

Most agent interfaces show a prompt and a final answer. The consequential part lies between them: scope changes, tool use, partial results and moments when a human should decide what happens next.

I am building Uxi to keep those decisions visible and editable inside the work itself.

• Represent an agent run as a workflow with steps, tools, checkpoints and failure behaviour, rather than a single conversation.
• Pause at consequential decisions so the user can approve, narrow scope, ask for evidence or redirect the next action.
• Attach the decision to a specific proposal revision, and keep observable actions and evidence in the project trace.
• Use replay and evaluation concepts to compare intended behaviour with what the agent did. This remains an active product direction.

I defined the interaction model and built parts of the Review Hub and workflow canvas in React and TypeScript, alongside local session and checkpoint plumbing. The canvas exposes tool access, memory, human pause points and failure behaviour as product decisions a user can inspect.

The working implementation is an independent product in development. Its review and evaluation approach is being tested; this case does not claim enterprise deployment or measured customer adoption.

An agent’s capability is useful only when people can understand its scope, intervene at the right time and inspect evidence after it acts. Uxi is my way of prototyping that experience end to end: product definition, UX, code and evaluation questions in one loop.