LanternOpsThe Operator's LogManaged AI Teams
LanternOps Book a diagnosis →
Entry 008 July 27, 2026

Week seven: half a week, on purpose

← The Operator's Log

Fleet telemetry · 2026-W30counted, not estimated
Leverage13.6×fleet hours per operator hour
Fleet hours257agents working
Operator hours19directing & gating
Shipped161artifacts out the door
Fuel$231tokens & tools
How these are counted →

Both numbers dropped by about half. 257 fleet hours, 19 of mine. The ratio barely moved, which is the part worth noticing: when I step back, the fleet steps back with me. It doesn’t run on its own for a week and hand me a pile. The engine is still gated on the operator, and the leverage number is a property of the pair, not of the machine.

I’m fine with that this week. I’d be less fine if it were still true in six months.

What shipped

Breeze. Four releases, v0.99.0 through v0.102.0. The home dashboard was rebuilt around accurate counts and honest failure states, because a dashboard that says “all clear” when it couldn’t reach the database is worse than no dashboard. Tier-3 AI actions can be self-approved with Touch ID, inline, and the tiers themselves were recalibrated against the real guardrail tables instead of a doc that had drifted from them. Admin MFA reset for the lost-authenticator case. Italian and Canadian French. A Huntress package in the built-in catalog. Windows patching learned to fall back to cumulative updates when a finding has no CVE tag, and to stop restart-storming healthy agents when the state file is locked.

Five production Sentry issues closed at the root, then four more. The partner-abuse detectors got two new signals. The July 20 security review was remediated the same week it was written.

Hive. One approval path. There were four: colony approvals, agent approvals, draft reviews, and trigger responses, each with its own table and its own way of failing. Now there’s one service, one inbox, and one fail-closed rule for when the service isn’t there. Agent memory went from a seam to on-by-default, with a kill switch. Discord DMs can create tasks.

Strata. Gate two passed, per-stage observability. Gate three, the unkeyed noninferiority test, failed on one falsifier, and the slice is paused there. The failure is recorded in the repo the same as the passes are. That’s the whole point of pre-registering the gate.

OliveTech. Workflow level. No client names.

What’s queued

Breeze starts a remediation program off the security reviews, in waves, with each wave recorded. Hive gets its settings out of environment variables and into the app, because a system you configure by editing a file on a server is not one you can hand to someone else. Strata sits at gate three until the machine is quiet enough to rerun it.

If you’re doing the math on what one person and a fleet can hold, there’s a door.