LanternOpsThe Operator's LogManaged AI Teams
LanternOps Book a diagnosis →
Entry 013 August 31, 2026

Week twelve: twenty to one

← The Operator's Log

Fleet telemetry · 2026-W35counted, not estimated
Leverage20.0×fleet hours per operator hour
Fleet hours548agents working
Operator hours27directing & gating
Shipped207artifacts out the door
Fuel$231tokens & tools
How these are counted →

Twenty to one. 548 fleet hours, 27 of mine. That’s the biggest fleet week I’ve counted and the widest ratio, and it happened without my hours changing at all. So the ceiling moved again, and this time it moved on the fleet side.

I know why. Most of those hours are orchestrated: a queue of issues, each one picked up by an agent in its own worktree, reviewed by another, merged by a third. My 27 hours are spent writing the specs at the top of that funnel and reading what comes out the bottom. The middle doesn’t need me and mostly doesn’t get me. Shipped is down, 207 against last week’s 260, which tells you the hours went into fewer, bigger things.

What shipped

Breeze. v0.108.0. The AI agents epic went from schema to running. Wave 3c is a headless runner in shadow mode: the agent does the work, records what it would have done, and does nothing. Wave 4 is act mode, with a closed manifest of operations that a rule could already perform, revalidated before execution, verified after, and always notified. Wave 5 is the policy-decide lane for unattended Tier-3, dark. Wave 6 is supervision: an execution-trace UI, an exposure readout, a per-org circuit breaker, and a kill state with an admin API. Then the first three pilots on top of it: alert verdicts, a ticket helpdesk in shadow, and a metric-anomaly trigger. And scheduled sweeps with a weekly narrative per org.

Every one of those waves is a PR you can read. An agent that acts on a client’s fleet is the thing this whole log exists to prove can be run responsibly, and the answer Breeze gives is: shadow first, closed manifest, exposure ledger, kill switch, a human at every tier that matters.

Also this week: the org lifecycle in five waves, including merge and archive. A worker split so the API container stops doing background work. A bring-your-own-LLM provider catalog with an egress enforcement layer. SSO linking that replaces a dead end with a confirm ceremony. A left-nav reorganization that split Settings from Operations, which should have happened a year ago.

Pressless. Import from Wix, Squarespace, GoDaddy, Webflow, and static HTML, with an honest analysis of what will and won’t come across. GitHub sync, phase one: a GitHub App, repo-scoped tokens, a manifest-only sync engine that detects divergence. Migration cost controls with estimates and a trial cap. The marketing site was trued up against what the product actually does, which is a sentence I’d like every company to be able to write.

Strata. Drain and stop. A service that can be told to stop and finishes what it admitted first.

Hive. Five commits, all Discord gate bookkeeping. Quiet on purpose; the work moved into Breeze.

OliveTech. Workflow level.

What’s queued

Breeze’s agents get their first real pilot on a real fleet, in shadow, and the log reports what the shadow would have done. The methodology page gets updated to name the orchestrators the fleet actually runs on now, because it still says Claude Code and Codex and that hasn’t been the whole truth since July. And the log itself catches up, because this entry is seven weeks late.

One more. The laptop all of this runs on has 24 gigabytes of memory, and at 548 fleet hours it became the wall. The replacement has 128 and arrives Monday. Next entry I’ll count the machines too, because the fleet is not one computer and I’ve never said so.

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