Memory, trust, and the front door

Four ships that close the loop between an autonomous team and the human it works for.

Checks that remember

A mission now carries working notes — the team's memory between checks. Each scheduled check reads them in its brief and rewrites them before finishing: open threads with how many consecutive checks they've been open, commitments with due dates, resolved items dropped. In live testing the aging emerged unprompted by the second check:

Catalyst Property Company — $80K (contractsent) [Check #2 open]

That's the difference between a team that rediscovers the world every morning and one that escalates what's been sitting too long.

The trust ladder's last rung

Approvals have trained the agents for a while (every decision becomes a learning rule). Now the payoff: workspace/<org>/trust.yaml can declare, per action, a confidence threshold above which proposals execute without waiting in the queue — fully audited, listed under Review's "Executed automatically," reversible by flipping enabled: false. The default is off, and no rules means nothing auto-executes, ever.

Alongside it, Review gained the proposals queue itself: every pending CRM update with its confidence chip, rationale, and the exact proposed change — Approve executes through the vaulted connector, Reject teaches.

Workflows can ask

A new ask step pauses a run with a question in Review — a textarea feeds the answer back as the step's output, interpolable downstream. The discovery follow-up now opens by asking for the call transcript, which means it's triggerable with nothing in hand (and ready for Zoom/Loom/Granola auto-fetch when those connectors land).

Briefings get a front door

The morning check now publishes its briefing (publish_briefing) to a first-party Briefings page — latest open, history collapsible. Open the app, read the brief, clear Review. That's the morning.