Browser Fleet
Fan browser work out across parallel subagents, each with its own isolated agent-browser session. Adopted from Greg's multi agent Chrome pattern 2026-07-23, collapsed onto our stack: agent-browser --session <name> gives per agent isolation natively (own cookies, tabs, refs), the Agent tool gives parallelism and result collection, so Greg's Chrome on ports scaffolding, per agent MCP configs, and chat.md coordination all disappear.
Tool Choice
- agent-browser (this skill) for everything unauthenticated or freshly authenticated: forms, scraping, sweeps, verification. Lightest context per action.
- Playwright MCP only when the task needs the dedicated authenticated Brave profile with Bitwarden (banking, YNAB, logged in web apps). Keep its footprint shrinking.
Execution
- Build the target list. Each work item is one URL plus the actions to take there (form fields, data to extract). Chunk targets across 2 to 5 agents; don't over provision (10 forms wants 3 agents, not 5).
- Spawn agents in parallel (single message, multiple Agent calls, or a Workflow pipeline for larger lists). Each agent's prompt includes:
- Its session name: every command runs as
agent-browser --session fleet-<n> <command>. Sessions are fully isolated from each other and from the default session.
- The recipe: run
agent-browser skills get core first; use snapshot -i refs and re-snapshot after every page change (refs go stale); --session on every single call; resolve the binary once with command -v agent-browser before any loop and use that absolute path inside loop bodies (the sandbox strips PATH there), so the recipe works wherever agent-browser is installed.
- Its target chunk with exact per target instructions and what to return (structured results, not transcripts).
- Cleanup:
agent-browser --session fleet-<n> close when its chunk is done.
- Collect and report. Agents return per target outcomes (submitted, blocked, captcha, data). Summarize wins, failures, and anything needing the user (captchas and logins bail, never guess credentials).
Guardrails
- State changing submissions stay gated. Any submission that creates records or sends messages on someone else's system is an outward action: show the user the payloads and the target list, and get an affirmative yes before the fleet fires. Without that approval the fleet stays read only. Same rule as email sends.
- Cap at 5 concurrent sessions; each is a real headless browser.
- A hung session gets closed and retried individually (
agent-browser --session fleet-<n> close). Escalate to pkill -f agent-browser only when the daemon itself is unresponsive, knowing it resets every session, not just the fleet.
Proven Use Cases
- Outreach contact forms across a prospect list (the July digital play sends).
- Lodging availability sweeps across booking sites for
assist:plan-trip.
- Parallel JS rendered scraping where
read alone isn't enough.
1---2name: browser-fleet3description: Parallelize real browser work across 2 to 5 agents, each driving its own isolated agent-browser session. Use when the same browser task repeats across many targets, like submitting contact forms across a prospect list, sweeping lodging sites for a trip, or scraping JS rendered pages. Triggers on "browser fleet", "parallel browsers", "fan out the browser", "submit these forms in parallel", or /browser-fleet. Not for single site sessions (plain agent-browser) or authenticated Brave profile flows (Playwright MCP).4---56# Browser Fleet78Fan browser work out across parallel subagents, each with its own isolated `agent-browser` session. Adopted from Greg's multi agent Chrome pattern 2026-07-23, collapsed onto our stack: `agent-browser --session <name>` gives per agent isolation natively (own cookies, tabs, refs), the Agent tool gives parallelism and result collection, so Greg's Chrome on ports scaffolding, per agent MCP configs, and chat.md coordination all disappear.910## Tool Choice1112- **agent-browser** (this skill) for everything unauthenticated or freshly authenticated: forms, scraping, sweeps, verification. Lightest context per action.13- **Playwright MCP** only when the task needs the dedicated authenticated Brave profile with Bitwarden (banking, YNAB, logged in web apps). Keep its footprint shrinking.1415## Execution16171. **Build the target list.** Each work item is one URL plus the actions to take there (form fields, data to extract). Chunk targets across 2 to 5 agents; don't over provision (10 forms wants 3 agents, not 5).182. **Spawn agents in parallel** (single message, multiple Agent calls, or a Workflow pipeline for larger lists). Each agent's prompt includes:19 - Its session name: every command runs as `agent-browser --session fleet-<n> <command>`. Sessions are fully isolated from each other and from the default session.20 - The recipe: run `agent-browser skills get core` first; use `snapshot -i` refs and re-snapshot after every page change (refs go stale); `--session` on every single call; resolve the binary once with `command -v agent-browser` before any loop and use that absolute path inside loop bodies (the sandbox strips PATH there), so the recipe works wherever agent-browser is installed.21 - Its target chunk with exact per target instructions and what to return (structured results, not transcripts).22 - Cleanup: `agent-browser --session fleet-<n> close` when its chunk is done.233. **Collect and report.** Agents return per target outcomes (submitted, blocked, captcha, data). Summarize wins, failures, and anything needing the user (captchas and logins bail, never guess credentials).2425## Guardrails2627- **State changing submissions stay gated.** Any submission that creates records or sends messages on someone else's system is an outward action: show the user the payloads and the target list, and get an affirmative yes before the fleet fires. Without that approval the fleet stays read only. Same rule as email sends.28- Cap at 5 concurrent sessions; each is a real headless browser.29- A hung session gets closed and retried individually (`agent-browser --session fleet-<n> close`). Escalate to `pkill -f agent-browser` only when the daemon itself is unresponsive, knowing it resets every session, not just the fleet.3031## Proven Use Cases3233- Outreach contact forms across a prospect list (the July digital play sends).34- Lodging availability sweeps across booking sites for `assist:plan-trip`.35- Parallel JS rendered scraping where `read` alone isn't enough.
mattforni