B-03 · in design
Agent Review Console
An approvals inbox built to shrink: agents earn autonomy one human yes at a time.
01 the number that stalled
Teams are shipping agents into production faster than they can supervise them. The share of organizations requiring human review before high-risk AI actions fell from 40% to 25% in six months, while full autonomy more than doubled (JumpCloud IT Trends Report, 2026). I run agentic workflows every day, and the difference between useful and dangerous is a human gate cheap enough that nobody skips it.
02 the system, specced
Full autonomy is a level an agent earns, never a setting you turn on. In the console, every agent action starts behind a human yes: each request arrives as a plain question, with the reasons written in sentences and the actual email shown exactly as the recipient would read it, and a decision takes one click. Every yes teaches the console. Approve a pattern enough times and it graduates to autopilot, so the inbox is built to shrink: in the month on the screens, 160 of 214 tasks ran on their own, and a person handled the 48 calls that needed judgment.
An operations or finance owner runs it day to day; the dev lead steps in only on escalation. Autonomy stays written down and reversible: every autopilot rule carries its provenance, its caps, and an undo, and every decision lands in a plain-language history that answers who authorized what. Risk grading runs live via the Anthropic API. The design descends from the stage-gate system I run in my own application infrastructure, where an agent drafts and a human gate decides what ships.
03 product screens
04 specced / scope
specced
- ✓Built to shrink: every approval teaches the console, and routine work graduates to autopilot, on the record and reversible.
- ✓Nothing acts without a yes: agents propose, people decide.
- ✓Every request reads as a plain question: what it does, why now, what it touches.
- ✓The artifact shows as the recipient will see it, an email as an email.
- ✓Escalation is one click: the odd case goes to the dev lead with its context.
- ✓A team view: every agent, its job, its leash, and the trust it has earned.
- ✓Risk grading generates live via the Anthropic API.
scope
- →No bulk approve, by design: the gate is the product.
- →One team in v1. Multi-tenant review is a later thesis.
- →Connections start simple: one system in, one system out. Deep integrations come after the loop proves itself.
status in design · concept and screens first, code after