Notes, 8 October 2026

Jeffrey Emanuel’s Agent Flywheel, against our loop

Jeffrey Emanuel publishes a method for running many coding agents at once, the Agent Flywheel. We read it at its source and held each part against our own loop and its ledger. Two of his practices were running here two days later, and several we left out, each for a reason.

His method, briefly

Plan at length with several frontier models, merge their plans, and refine the result over several rounds. Turn the plan into small tasks with dependencies and polish those again before anyone builds; in his words, “check your beads N times, implement once”. Then run five to fifteen agents in parallel on one branch, coordinated by messages and file reservations. Each agent rereads its own diff with fresh eyes, others review it in turn, scanners run before commits, and a person tends the swarm every ten to thirty minutes.

What we took

Fresh eyes, as it is: every build now ends with the builder rereading its whole diff. In the kit on 10 October.
Polishing before building, changed: a second model family reads every medium or large card Claude writes, and the card waits until that reader says it is ready, unless you waive the read. In the kit on 10 October.
The person at the plan, changed: his method keeps no approval gate on the diff. Here, cards you pre-approve, in repositories you mark low-risk, may merge overnight through a merge pass that ships switched off, and changes to the loop itself always wait for you. Built from 11 October.

What we left out, for now

Agent swarms on one branch, with Agent Mail and bv: they assume someone tending agents every ten to thirty minutes. Our queue is short, so one build slot does the work. We will look again when the queue stays backed up.
Session search and memory tools, cass and cm: the habit is right, and our ledger already keeps every attempt, review and receipt.
A one-command bootstrap, ACFS: the kit already provisions with a bootstrap script and an Ansible play, so there was nothing to take.

What it confirmed

One operating manual, reread after the context is compacted.
Command-line tools that print data for agents and diagnostics for people.
Building blocks that work the same under any agent.
A second approval before anything dangerous runs; here it is yours.

Still open

A guard against destructive shell commands. Nothing is in the kit for it yet.

The tags it produced

All the tags.

Sources

A credit names where an idea came from. None of these people has endorsed the kit.