---
title: Features
description: Why UseAgent. The agent contract, the trust lane, the reliability story, replaceable engines, the workspace, the self-improvement loop, and the product experience, each as a deep dive.
sidebar:
  label: Why UseAgent
  order: 1
---

UseAgent's pitch fits in one sentence: **a prompt becomes a durable, replayable
run**, executed by the coding engine of your choice inside an isolated Linux
workstation, rendered live from one canonical event log, and finished with
everything worth keeping captured for review.

This section unpacks that sentence one commitment at a time. Each page is a deep
dive into a property the platform actually enforces, written from the shipped
code rather than from aspiration.

## The deep dives

<CardGroup cols={2}>
  <Card title="The agent contract" href="/features/agent-contract" icon="file-check">
    The flagship spec. One provider-neutral event grammar with a versioned
    envelope, a negotiated capability map, and delivery guarantees every engine
    must satisfy. This is what makes engines swappable.
  </Card>
  <Card title="Trust and control" href="/features/trust-and-control" icon="shield-check">
    The sandbox is untrusted by design. Capability tokens, a separate gateway
    process on a restricted database role, one-shot human approvals, and
    write-only encrypted secrets.
  </Card>
  <Card title="Reliability" href="/features/reliability" icon="rotate-ccw">
    Runs survive backend restarts, reconnects replay exactly what live viewers
    saw, retries are idempotent, and releases must pass real engine journeys
    before they ship.
  </Card>
  <Card title="Engines" href="/features/engines" icon="plug">
    Codex, Claude, and OpenCode behind a neutral adapter. The inner loop is
    rented; orchestration, history, approvals, and rendering stay yours.
  </Card>
  <Card title="The workspace" href="/features/workspace" icon="monitor">
    A real Linux workstation per thread: repositories, terminal, a visible
    desktop with browser automation and recording, artifacts, and editable
    workpieces.
  </Card>
  <Card title="Self-improvement" href="/features/self-improvement" icon="brain">
    Runs recall what the organization knows and propose what they learned.
    Every publish passes a human gate, and every recall is auditable.
  </Card>
  <Card title="Product experience" href="/features/product-experience" icon="layout-grid">
    A native React surface owned end to end: one timeline grammar for every
    engine, a side rail of live panes, a command palette, themes, and Slack as
    a first-class channel.
  </Card>
</CardGroup>

## How these pages are written

:::note
**Grounding discipline.** Every claim in this section is verified against the
code it describes, and file paths are cited where the mechanism lives. There
are no invented benchmarks and no aspirational numbers: where a count appears
(30 canonical event kinds, 20 capability flags, a 22-case release parity
matrix), it is counted from the source. Where a capability is narrower than
the ambition, the page says so.
:::

If you want the neutral reference instead of the argument, the same ground is
covered dryly in [Concepts](/concepts), [Architecture](/architecture), and
[Product](/product).