Browser and desktop agents

Browser and desktop agents perform work in the same software people use.

They need more than screenshots and clicks.

They need identity, durable sessions, recoverable computers, bounded permissions, and evidence of each material action.

Computers available now Bring your agent model

Choose the right execution surface

Surface	Use it for	Avoid it when
API or MCP tool	Structured, supported application actions	The required action is not exposed
Sandbox browser automation	Disposable, reproducible web tasks	Login identity must persist for months
Persistent computer	Authenticated websites and desktop applications	A reliable API already exists
Human handoff	Ambiguous, sensitive, or irreversible decisions	The action is routine and bounded

Prefer structured APIs when they preserve the same business meaning.

Use a computer when the real workflow exists only in a website or graphical application.

Product patterns

Back-office operator

Works through portals, spreadsheets, inboxes, and administrative systems on behalf of an operations team.

Desktop RPA product

Automates legacy applications while allowing a reviewer to accept, reject, or revise individual steps.

Authenticated research agent

Uses paid databases and customer-owned accounts without asking the user to log in for every run.

Design production agent

Operates browser-based creative tools, exports artifacts, and presents a review queue.

QA and testing agent

Navigates a real product, captures visual evidence, reports regressions, and verifies fixes.

Procurement agent

Collects quotes, enters order information, and pauses before a purchase or contract action.

Identity and session model

flowchart LR
  User["Customer or team"] --> Identity["Agent identity"]
  Identity --> Computer["Persistent computer"]
  Computer --> Browser["Browser profile"]
  Browser --> Systems["Customer systems"]
  Computer --> Vault["Scoped secrets"]
  Computer --> Evidence["Screenshots and action log"]
  Evidence --> Review["Human review"]

Give each customer, team, or regulated workflow an explicit computer ownership boundary.

Do not silently share browser profiles between unrelated customers.

Store credentials through scoped secret bindings rather than source files or prompts.

Recovery behavior

A computer may pause, migrate, or restart.

The product should distinguish computer health from application readiness.

After recovery, verify the desktop process, browser profile, required applications, network route, and target session before returning work to the agent.

Pause is not deletion

A paused computer should retain its durable state. The control plane should report whether it can resume and what was verified after resumption.

Human control

Require approval for purchases, destructive actions, permission changes, external communications, and high-impact account updates.

Show the proposed action, affected system, expected effect, supporting evidence, and rollback path.

Computer APIs Create, operate, embed, pause, and resume full desktop computers.