Back to blog
3 min read
aiai-agentsstartups

One Person Many Agents: Manager Pattern That Works

Nous ran 1,393 agents to cut 1M lines and Claude Projects added a coordinator on Sep 17. The manager pattern that lets one person run many agents safely.

Share
One Person Many Agents: Manager Pattern That Works
On this page

One manager, many workers, zero overlap. That pattern links two September stories: Nous Research refactoring one million lines with 1,393 subagents, and Anthropic relaunching Claude Projects with a coordinator over parallel threads on September 17. Here is how the pattern works in plain terms, and how to copy it without a team.

Takeaways

Nous result: 1,063,826 lines to 698,363, files over 5,000 lines from 37 to 6, for $19,300 plus follow ups near $25,000. Method that made width safe: 36 non overlapping groups, a manager that never edits code, byte for byte interface checks. Projects version: shared memory plus goals, one thread per branch, conflicts resolved like pull requests, cloud only at launch. Solo rule: split first, then fan out, then check mechanically. Review taste with humans, not with more agents.

Why did 1,393 agents not collapse into chaos?

Because the work partitioned. Teknium counted the Hermes repo, split code into 36 groups with no shared files, and gave each worker one group plus a skill file built from months of corrections. The orchestrator never edited code, it assigned, collected, and verified interfaces exactly. Peak width hit 218 workers from one desktop with Claude Fable 5.1 for inference, resuming once after an auth token expiry.

manager:   partition repo -> assign groups -> verify interfaces -> merge
workers:   clean one group -> keep public names -> fix suppress calls -> report
checks:    byte compare APIs -> community review -> post merge fixes

What review caught

Workers removed public names with no internal callers that external plugins used, and rewrote suppress calls at 65 sites in ways tests missed. Both were fixed before merge: partitioning limits blast radius, review still catches contract breaks.

How does Claude Projects map to the same idea?

1

One conversation on top

Talk to the coordinator, not each worker. The coordinator holds goals, memory, files, and artifacts shared across threads.

2

One branch per thread

Each thread runs as a cloud session on its own repo copy. Overlaps surface as merge conflicts, a workflow developers already know how to resolve.

3

Spawn narrow subagents inside threads

Threads can add loops and structured workflows for parallel pieces. Keep each subagent on one assignment with its own context.

4

Stay cloud aware at launch

Beta starts with select Pro and Max users, cloud threads only, local tools marked very soon. Keep private deps out until local support lands.

When should you keep width small?

When tasks share files, when success means taste, or when agents check each other by reading output. Eric Provencher warned September 16 that more than two reviewing agents burns tokens because agents double check homework. Matt Shumer argued the same day for one manager farming out work. Both fit Nous: one manager, wide execution, mechanical checks, no agent juries.

What should a solo setup look like?

What is the smallest fleet that works?

One manager session plus two to four workers on separate folders or branches, each with a written brief from a skill file. Add workers only after byte checks pass three runs in a row.

How do I measure if parallel work pays?

Track wall time, tokens per merged change, and rework rate. Nous cut lookup tokens per symbol from 2,218 to 993: measure reading cost plus merge cost, not just lines removed.

As of September 24, 2026: leverage comes from memory plus partition, not headcount. Write the skill file first, split so workers never touch the same file, and let the coordinator merge. Next, read why AI got cheap and fast to price that fleet before you scale it.

Questions, answered

What did the Nous 1,393 agent run prove?
Starting September 2, 2026, Hermes dispatched 1,393 subagents over 19 hours with 218 at peak, merging September 4: non test Python fell 34.4 percent for about $19,300, using 36 non overlapping groups plus byte checks.
What is Claude Projects in September 2026?
A September 17, 2026 relaunch inside Claude Code: each Project holds shared memory and goals, threads run as cloud sessions on branches, and a coordinator resolves overlaps as pull request conflicts.
How many agents should a solo builder run?
One manager plus a few workers on split tasks. Width helps only when work partitions cleanly and checks are mechanical: shared review of taste based work still burns tokens.
Share

Founding software engineer and curious tinkerer, writing about AI, systems, and the craft of shipping.