How two AI agents share one Chrome without colliding

If you run two AI agents against one browser, the first problem you hit is not intelligence. It is traffic. Both agents want a tab, both may want the same one, and neither knows what the other just did. One of the most common questions we heard after launch was a simple version of this: how do two agents not collide, or submit the same form twice?

This post explains how Beacon handles it. The short answer is that agents do not share a tab. Each one works in its own.

A browser tab is shared, mutable state. If two agents act on the same tab, each one reads the page and then acts on what it read. By the time the second action lands, the page may have changed because of the first agent. A click meant for one page ends up on another. A form gets filled by one agent and submitted by the other, or submitted twice.

Nothing about that needs a bug in either agent. It is simply what happens when two writers share one surface without any rule about who owns it.

Beacon gives each agent a lane. A lane is the agent's own tab group in Chrome, with its own colour. The agent opens its tabs inside that group and works only in those tabs.

Because the group is colour-coded, you can see who owns which tab at a glance. Two agents mean two groups. They can both be busy at the same time without ever pointing at the same tab.

Sometimes an agent will try anyway. It might hold a stale reference to a tab, or be asked to act on a page that belongs to someone else. In that case Beacon does not act. The agent gets a WRONG_LANE error instead.

That rule covers two cases:

• Another agent's lane: an agent cannot click, type or navigate in a tab that belongs to a different lane.
• Your tab: the tab you are using yourself is not the agent's to touch. In agent mode it stays untouched while the agent works in background tabs.

The point of an error, rather than a silent best effort, is that the agent finds out. It can look at the page again, open its own tab, or ask you. It does not quietly do the wrong thing.

It removes the usual cause. Two agents in one tab can both submit a form because they are both looking at it. With separate lanes, each agent has its own tab, so there is no shared form for both to submit. If you ask two agents to do the same job, they will each do it in their own tab, so the job itself is still yours to define sensibly. Lanes keep agents out of each other's way; they do not decide what you ask them to do.

The Agent Board in the side panel shows the running agents, lets you stop one, and lets you add another. When a job is finished, the agent closes the tabs it opened and removes its tab group. Your own tabs stay put.

On the free plan you can link 3 Chrome profiles to one account and run 5 agents at once in each. Extra jobs wait their turn instead of failing.

Install the extension, connect your MCP client, and ask for two things at once. The setup guide covers Claude Code, Cursor and any other MCP client. For the longer version of this idea, see how to run multiple AI agents in Chrome.