Autobrowse

Purpose

Autobrowse turns expensive browser exploration into durable operational memory. It runs a real browser task, studies the trace and diagnostics, iterates on the strategy, then graduates the reliable path into a reviewable skill artifact.

Routing

• Use when: the user wants a repeatable web workflow learned from live exploration, optimized for future agents, or packaged as a skill.
• Do not use when: the work is a one-off lookup, one-page scrape, generic research task, or frontend QA task.
• Outputs: strategy.md, iterations.md, draft-skill.md, and usually an approval-gated skill_proposal.
• Success criteria: the learned workflow is repeatable, evidence-backed, and ready for human review or future agent reuse.

Operating Loop

The user should not need to provide structured fields. Treat a plain request plus an optional link as enough input. Infer the objective from the request, infer the target site from any URL/domain in the request, default to 3 iterations, and default to creating a proposal.

1. Objective. Restate the user-visible task and target site. Define what counts as success and what actions would be irreversible.
2. Attempt. Use the browser normally. Prefer visible Browser V2 tooling when interacting with pages.
3. Diagnostics. Use browser_console, browser_network, browser_storage, browser_snapshot, and browser_evaluate when available. Use browser_trace_start and browser_trace_stop as supplemental diagnostics when the runtime exposes a readable trace summary. Capture only redacted, relevant evidence.
4. Study. Identify stalled steps, brittle selectors, hidden APIs, unnecessary clicks, rate limits, waits, auth boundaries, and deterministic shortcuts.
5. Strategy. Update strategy.md before the next iteration. The next attempt must read it first.
6. Iterate. Repeat up to the requested cap. Default to 3 iterations and never exceed 5 unless the user explicitly asks.
7. Converge. Stop early when the task succeeds and another pass produces no material improvement.
8. Graduate. Write draft-skill.md. If safe and useful, create a skill_proposal so the user can approve the new skill.

Safety Rules

• Do not bypass captchas, paywalls, login restrictions, authorization boundaries, or consent gates.
• Pause before purchases, submissions, account changes, messages, destructive actions, or anything with external side effects.
• Redact credentials, tokens, cookies, private profile fields, and personal data not required for the workflow.
• Prefer public, documented, or naturally exposed endpoints. If an endpoint is undocumented, describe how it was observed and the risk that it may change.

Evidence To Capture

• Start URL, final URL, and success state.
• Browser actions taken, especially unnecessary or removed steps.
• Stable selectors, ARIA labels, form field names, or URL patterns.
• Network endpoints, request parameters, response shape, and headers needed for safe replay.
• Console errors, auth/permission problems, rate limits, anti-bot signals, and timing waits.
• Tool/action counts per iteration when practical.

Graduation Requirements

A graduated skill must include:

• Purpose and clear routing boundaries.
• Inputs and required parameters.
• Primary workflow, with deterministic paths first and browser fallbacks second.
• Validation checks that prove the workflow succeeded.
• Site-specific gotchas and freshness date.
• Safety boundaries and side-effect rules.
• Helper script names or direct fetch examples when they reduce browser work.

Use skill_proposal with action create by default. Use draft-only when the workflow is too fragile, too sensitive, or still missing validation evidence.

An Autobrowse run is not complete until strategy.md, iterations.md, and draft-skill.md exist in the run directory. If proposal creation fails, record the exact failure in iterations.md and keep the draft skill reviewable.