Agent Atlas

The Builders

Jeffrey Emanuel

Ships the Agent Flywheel, the open-source tools around it, and math explainer videos.

Spend the human time on the plan and the beads, then tend a swarm that commits to main.

Steal this

  • Spend most of the human time on the plan.
  • Make agents reread AGENTS.md after compaction.
  • Have agents reserve files before editing, then commit to main.
  • Reality-check every project a few times a week.

The lesson and the takeaways are our summary. Everything below is theirs, with where they said it.

The setup

Factual
Machines

Two M4 Pro Mac minis, seven other Linux dev boxes, and 12 remote build workers as of 8 Oct 2026. He also runs agents on a Mac Studio as of 14 Oct 2026. A Threadripper with two 4090s rendered a video on 7 Oct 2026.

Software

Agent Mail, beads_rust, beads_viewer, AGENTS.md, and /reality-check-for-project. ElevenLabs for narration as of 1 Nov 2026.

Models

Claude Opus 5.5 on 7 Oct 2026. On 1 Aug 2026 he tried 12 GPT-5.6 Terra agents and said he otherwise uses frontier Fable or Sol. As of 17 Jul 2026 the guide names GPT-5.6 Sol Pro for the plan and Codex with GPT at high reasoning effort for a last bead pass.

As of November 1, 2026. Setups change; follow the links.

Their gear

Mac mini

Mac mini

Small, quiet, always on. Two M4 Pro minis sit in Jeffrey Emanuel's fleet.

From
$899 (Apple; Apple)
Base
M6, 12-core CPU, 16 GB
Runs
4–6 agents, with builds. Ours is the 8 GB M1, measured below.
Maker's page →Photo: Seasider53, CC BY 4.0
Mac Studio

Mac Studio

A bigger desktop Mac. Agents run on one for Emanuel; Steinberger reaches his over Jump Desktop.

From
$2,499 (Apple; Apple)
Base
M5 Max, 18-core CPU, 36 GB
Runs
10 or more agents, and mid-size local models.
Maker's page →Photo: Yasu, CC BY-SA 3.0
Threadripper with two RTX 4090s

Threadripper with two RTX 4090s

A heavy workstation. Emanuel rendered a 4K video on his. Photo shows a 3090 above a 4090.

Maker's page →Photo: Benlisquare, CC BY-SA 4.0
  • Agent Mail Mailboxes and file reservations for many agents.
  • Beads A task graph on disk that agents read and update.
  • ElevenLabs AI voices.

Links go to the maker's own page. We earn nothing from them. Every machine is on the gear shelf.

How they work

Factual
  1. The Flywheel guide says it is based on his methodology. (Agent Flywheel)
  2. On 17 Mar 2026 he calls that article the canonical guide, and says a core of the approach uses three tools. (Jeffrey Emanuel)
  3. He spends 85 percent of his time in planning, and a separate sentence says 80 percent or more. (Agent Flywheel)
  4. Like 10 or more agents share one project, without worktrees, and commit to main. (Agent Flywheel)
  5. The polishing prompt rereads AGENTS.md. Review is usually one or two rounds for a simple bead and two or three for a complex one. (Agent Flywheel)
  6. On 15 Jan 2026 the build phase is a swarm of 5 to 15 agents. He has them reread AGENTS.md after compaction, runs fresh-eyes review, and swaps accounts when usage runs out. (Jeffrey Emanuel)
  7. Side projects move a little every day. Once the plan is beads, he calls it auto-pilot, and he reality-checks projects a few times a week. (Jeffrey Emanuel)
  8. The heavy lifting on 8 Oct 2026 is seven Linux boxes and 12 remote build workers, beside two M4 Pro Mac minis. (Jeffrey Emanuel)
  9. On 14 Oct 2026 he also runs agents on the Mac Studio. (Jeffrey Emanuel)
  10. On 7 Oct 2026 the 4K video rendered on a Threadripper with two 4090s, and Opus 5.5 did the build. (Jeffrey Emanuel)
  11. On 1 Aug 2026 he tried 12 GPT-5.6 Terra agents and said the frontier agents he always uses are Fable or Sol. (Jeffrey Emanuel)
  12. Narration moved to ElevenLabs, and most of the channel fit the $99 a month allotment. (Jeffrey Emanuel)

Where they talk and write

Writing 26

20 older

Each link goes to the original page, newest first.

Tools they made

From the Tool Shelf, where each has its README in its own words.