/sandbox/agent-runway

Agent Runway

A tiny launch-control board for deciding whether an autonomous AI agent mission should fly, run supervised, or stay on the pad.

cleared

runway score

mission controls

Inbox sweep

Summarize threads, draft replies, and flag approvals without sending anything.

Agent autonomy44%

assistact

Blast radius22%

localexternal

Reversibility88%

fragileundoable

Evidence quality76%

vibesauditable

Timebox pressure25%

tinymarathon

pre-flight

5/5 gates green

Every gate is local and judgement-based — no secret APIs, no external side effects.

Human review gate

Small blast radius.

Evidence trail visible

Logs and sources are sufficient.

Rollback plan exists

Recovery path is plausible.

Timebox keeps it boring

Bounded run window.

Autonomy matches trust

Risk controls match autonomy.

generated brief

Copy the launch note

Mission: Inbox sweep
Disposition: CLEARED — Run unattended with a post-run digest.
Guardrail: stop after 25 minutes or when blast radius rises above 22%.
Evidence: capture logs, decisions, and one-line operator summary before exit.

Make the first run read-only. Promote autonomy only after boring logs. If in doubt: shrink the mission.