Agent Steering

Correct an agent while it works, without stopping it and starting again.

PROGet Agent Steering and 168 other Pro components with pro lifetime — Get Pro

Category
AI Agent
Access
Pro
Source
410 lines · one file
States
6 drawn
Added
21 Oct 2026

The decision inside it

A note waits for the next step boundary

A correction never lands halfway through a tool call, which is how agents write half a file. The queue says which step will be the first to see it ("Applies from: Draft the reply"), so the timing is visible before it happens.

When to use it

Use Agent Steering beside a running agent so a person can correct it without stopping it: notes queue for the next step boundary, can be edited or pulled until picked up, and Stop after this step pauses cleanly. Your server marks a note picked up by setting pickedUpAt.

Installation

How you get it

Paste this into Claude Code, Cursor, Codex or any coding agent. It installs the file and uses it as it is, instead of writing a new one.

Use Liro's Agent Steering for this.
Install it: npx shadcn@latest add @liro/agent-steering
(the @liro registry is set up in components.json, with LIRO_KEY in .env.local)
Import it from @/components/ui/agent-steering.
Do not rewrite or restyle it. Pass my data through its props
and keep its built-in states.

For the full brief with every prop and the mistakes assistants usually make, use Copy prompt above. Pro components install through the @liro registry with your CLI token: see the Namespaced tab.

Usage

components/steer-run.tsx
import { AgentSteering, type SteerNote, type SteerStep } from "@/components/ui/agent-steering";

const steps: SteerStep[] = [
  { id: "read", label: "Read the renewal thread", status: "done", detail: "3 emails" },
  { id: "draft", label: "Draft the reply", status: "running" },
  { id: "send", label: "Send it to Olivia", status: "pending" },
];

export function SteerRun({ notes, api }: {
  notes: SteerNote[];
  api: { steer: (text: string) => Promise<void>; interrupt: () => void; resume: () => void };
}) {
  return (
    <AgentSteering
      agent="Renewal agent"
      steps={steps}
      status="running"
      notes={notes}
      onSteer={api.steer}
      onInterrupt={api.interrupt}
      onResume={api.resume}
    />
  );
}

API reference

“Copy prompt” above puts a full brief on your clipboard, install command, dependencies, props, and the mistakes an assistant usually makes with this component.

Props

Read from the component’s own TypeScript interface when the site is built, so this table cannot drift from the file you install.

TypeNotes
stepsSteerStep[] Yes–
statusSteerRunStatus Yes–
notesSteerNote[] NoNotes sent so far: queued ones have no `pickedUpAt`.
titlestring No–
agentstring No–
onSteer(text: string) => void | Promise<void> NoSend a note. May return a promise; the box waits for it.
onEditNote(id: string, text: string) => void No–
onRemoveNote(id: string) => void No–
onInterrupt() => void NoStop after the current step returns.
onResume() => void No–
placeholderstring No–
lookLook NoLayered sets it in a tray, the heading above the card and the status in the strip below. Flat is one card.

Dependencies

npm packages

clsxframer-motionlucide-reacttailwind-merge

Liro foundations, installed alongside it

liro-themebutton

States

All of them are drawn by the component. None is left for you to design.

  • Running
  • Note queued
  • Picked up
  • Stopping
  • Paused
  • Done

Accessibility

  • Keyboard reachable throughout, in visual order
  • Focus trapped in overlays, Escape closes them
  • Focus returns to whatever opened it
  • Live regions announce what changed, not that something did
  • Honours prefers-reduced-motion

Responsive behavior

  • Sizes to its container, not the viewport
  • Usable from 320px up
  • Pointer-based interaction, so trackpad and touch both work
  • No hover-only affordances

Used in a screenAll screens

ScreenPro

Multi-agent Chat

A chat with the agent running a job, the agents it handed work to, and the brief, side by side.