Parallel Agents (Swarm)

AI Browser Agent

Run one task as several agents at once — split a batch for speed, or collaborate on a single shared page.

On a Cloud Browser profile the Parallel agents counter is always available: in the Run settings dialog on the start screen, and in the chat's status bar until you send the first message. For a local Chrome profile, first turn on Parallel agents (swarm) in Settings → AI Browser Agent. You choose how many agents a task runs as — the counter stops at your plan's maximum, and on the Cloud Browser each agent also uses one of your plan's concurrent cloud browsers. You don't pick a mode: the system reads the task and runs it as Independentagents, as a Collaborate on one shared page team, or as a single agent when it isn't safe to split — and the two parallel modes behave very differently.

Parallel agents apply to a chat's first message only. A follow-up continues that chat's own browser and context, so start a new chat to run another task in parallel.

Parallel agents are also available on schedules — the Schedule-an-Agent dialog carries the same count, and every scheduled run fans out server-side and merges the results — and on the Visual Studio browser-agent node (the “Parallel Agents” field, for cloud/scheduled/API executions).

The Parallel agents counter in the run dialog, set to 3 agents — The run dialog's Parallel agents counter — pick how many agents a new chat's first cloud run fans out across (1 = single agent).

Two ways a swarm runs

Independent — split a batch (most common)

The work is divided across agents and each does its own slice in parallel.

Best for batch work: “fill these 25 forms”, “check these 30 product pages”, “visit each of these links and grab the price”. Each agent takes part of the list and works on its own. Nothing is shared between them; their results are combined at the end.

Each agent can run in its own Chrome profile / window for true parallelism (see below).

Collaborate on one shared (real-time) page

All agents work on the SAME live page and see each other's changes.

Chosen only when the target is a real-time shared page — a live board, a collaborative document, a multiplayer canvas — where every participant sees the others' edits instantly. The task is split by section/area so the agents don't collide (e.g. agent 1 takes one column, agent 2 another), and a short teamwork protocol is added so they coordinate.

Each agent still runs in its own Chrome profile / window — they are separate participants that meet on the same URL, exactly like several people opening the same shared link.

Not for ordinary sites

If the page isn't a real-time shared surface, collaborate mode gives no benefit — the agents can't see each other's work — so the system uses Independent mode or a single agent instead.

Per-agent Chrome profiles (local)

Local Chrome

Give each agent a different Chrome profile and they run in separate windows simultaneously — true parallelism. Leave an agent set to “Same window as Agent 1” and it shares one window using tabs (simpler, but slower, since the window can only show one tab at a time).

Need more profiles? Use the Profile menu → New Chrome Profile. They then appear in the per-agent picker. (This applies to both Independent and Collaborate mode.)

Cloud browser

On a cloud profile, each agent gets its own cloud browser automatically — there are no local profiles to pick. Your plan's concurrent cloud-browser limit caps how many run at once.

When does each one run?

One agent — a single task. Just leave the count at 1.

Independent (2+) — a list/batch of similar jobs to finish faster.

Collaborate (2+) — a single real-time shared page that several agents organise together.

Cost

Parallel agents split the work rather than repeat it, so they save time more than AI credits. Each agent runs in its own browser, though, so cloud-browser time adds up per agent. Start small (2–3 agents) to gauge the cost and speed before scaling up.