Agents Contract
Public readiness contract for prompt-driven browser agents.
Public entry points must stay traceable, bounded, and production-safe.
Runs need logs, browser evidence, limits, replay, and API semantics.
Current workbenches run on managed sessions and remain the base.
Available Paths
Public agent entry points are governed by the readiness contract. Use the available production surfaces today:
Use the Browser workbench. It already has real sessions, live view, logs, network events, and replay.
Use Fetch for URL-first text, HTML, and screenshot output.
Use Crawl for max pages, same-site link following, and include/exclude boundaries.
Public Launch Contract
Before agents become public, the run object must be traceable, steps must be explainable, limits must be visible before execution, and evidence plus usage must connect back to the same run. Every run needs a session id, run id, status, started/ended timestamps, terminal state, action timeline, and failure reasons.
Before a run starts, users must see the task prompt, executor, model or provider, timeout, session retention, profile access, permissions, and recording mode. These are run inputs, not dashboard-private state.
Results and Handoff
The result view and public API must describe the same run object: final answer or terminal failure, step timeline, browser evidence, output artifacts, and usage semantics. SDK, cURL, team API keys, and public OAuth should all return the same run structure rather than a dashboard-only result.
Readiness comes down to four checks: failures are diagnosable, risk is controllable, results are reproducible, and usage is explainable. Agents must not bypass the session model, hide browser actions, or launch unbounded execution by default; they should compose Browser, Fetch, and Crawl rather than replace them.