LanternOpsThe Operator's LogManaged AI Teams
LanternOps Book a diagnosis →
Entry 007 July 20, 2026

Week six: the idea I stopped looking at

← The Operator's Log

Fleet telemetry · 2026-W29counted, not estimated
Leverage12.5×fleet hours per operator hour
Fleet hours493agents working
Operator hours39directing & gating
Shipped387artifacts out the door
Fuel$231tokens & tools
How these are counted →

Last week I wrote that Breeze was loud and everything else was quiet, and that I was letting volume decide for me. This week the quiet thing got two hundred commits. I don’t think writing it down caused that. I think writing it down made it impossible to keep not doing it.

The numbers: 493 fleet hours against 39 of mine, twelve and a half to one. My hours went back up from 28 to 39. That’s not the fleet getting slower. That’s me sitting in more loops, most of them in Strata, where nobody has built the thing before and the review can’t be delegated yet.

What shipped

Breeze. Eight tagged releases in seven days, v0.95.0 through v0.98.1, all on GitHub. That number is not entirely good news; see the miss.

The substance underneath: a versioned runtime extension contract, so what gets added to Breeze later goes through a signed bundle and a testkit instead of a side door. Tier-3 approvals became durable action intents, which means an AI action that needs a human doesn’t evaporate if the browser closes. The chat learned to answer over the estate with citations instead of confidence. Quotes got a document-styled editor with a thirty-second undo-send window, which is the kind of feature you only build after you’ve sent the wrong one. Linux remote desktop landed in pure Go. Spanish, French, and German locales went in. TD SYNNEX distributor pricing became a live lookup and a nightly ingest.

Two more security waves closed: the trusted client-IP boundary now fails closed, and an API key can no longer outlive the authority of the person who created it.

Hive. The counter reads Codex now. That was the queued item from last week and it’s done, along with the transcript retention fix, so the number on this strip is the first one computed from every source the fleet actually runs on. The rest of Hive’s week was live registry work: one shared registry across concurrent runs instead of one per run.

Pressless. A monthly SEO report pipeline, end to end: collectors, an arithmetic core the AI is not allowed to touch, a verifier, a renderer, and an admin surface. Google Business Profile connect, phase one, admin-only. And a Sequenzy email marketing integration. The pattern is the same one I use everywhere: the model interprets, the math is deterministic.

Strata. The one I said was untouched. It now has a Rust coordination kernel, durable publication, crash recovery, and a live multi-agent comparison against the way everyone does this today, which is one git worktree per agent and a merge step afterward. In every evaluable scenario the Strata arm won on both cost and time to a shared green codebase. Three runs each, one model, one machine. That’s a direction, not a proof, and the write-up says so. It’s on npm.

OliveTech. Workflow level. Same triage stack, same gate.

What’s queued

Breeze slows the release cadence down and puts the migration guard in front of it. Strata gets its second gate, per-stage observability, before I let myself believe the first one. The Field Notes LinkedIn series is still on this list, which is now three weeks running, and I’m leaving it here on purpose.

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