ForkLab Workbench

Enterprise agent workflow (Preview)

Prompt → generate patch → BrowserPod run → proof report

AI-generated code is untrusted until BrowserPod verifies it.

Agent task

Secure run request

Ready

Agent activity

Timeline

Issue selected

Built-in repo context is staged.

done
Secure sandbox run

BrowserPod boots and receives source files.

waiting
Failure evidence

The tenant access-control test fails first.

waiting
Patch proposal

A provider returns a constrained full-file patch.

waiting
Human approval gate

ForkLab waits before writing AI output.

waiting
BrowserPod verification

The approved patch is executed and tested.

waiting
Verified fix accepted

The result is ready for handoff.

waiting

Verification policy

Trust gate

First test runnot-run
Patch proposalwaiting
Human approvalrequired
Final BrowserPod runnot-run
/workbench · secure-agent-run
BrowserPod terminal
Ready. Select SEC-101 and launch a secure agent run.

Patch proposal

Waiting for secure run

Launch SEC-101 to generate a patch proposal. No patch is written until approval.
Proof report: pending

The proof report completes only after the approved patch passes inside BrowserPod.

Failure observedWaiting
Patch approvedWaiting
Passing test observedWaiting

What is real vs preview?

Honest demo boundary

Real

SEC-101 writes files and runs Node tests in BrowserPod.

Real

/api/agent/plan-patch keeps Gemini and Groq keys server-side.

Real

Fallback mode is deterministic and still requires approval.

Preview

GitHub import, queued UI-204, and PR creation are not live yet.