Agent Swarm Mode

What it is

Agent Swarm Mode is Next Chapter's multi-agent intelligence system. When Ewing has a high-stakes question — who are the 80 most likely buyers for HR.com? or should we consolidate 16 repos into one? — he doesn't ask one AI. He deploys a team of 11 specialized agents on the problem simultaneously.

Each agent has a different lens: one watches the numbers, one hunts for the people behind the deals, one asks "what does the human on the other side feel?", one makes sure the narrative is coherent, and one independent auditor checks the final answer without ever seeing how it was made. The output is one clear recommendation — with any unresolved disagreements named and cited.

The roster (11 agents)

These 11 are part of a larger fleet. Browse all 56 skills →

How it works — the 6 phases

Every swarm run follows the same contract. Skipping a phase is not allowed — audit-quality will catch it.

What it's done — 9 runs

Every run is archived and bookmarkable. Click any title to read the full output, dissent, journals, and signals.

Where results live

Everything is public on chapter.guide. Bookmark any of these URLs.

How to call it

In Claude Code, type the slash command:

You can also specify a subset to skip agents you don't need:

Presets

What happens next

1. Conductor locks a run number and opens a notebook entry
2. All agents receive your brief simultaneously (Phase 1)
3. They cross-read and challenge each other (Phase 2)
4. audit-quality reviews the consolidated answer cold (Phase 4)
5. tech-translator strips jargon from the final output (Phase 5)
6. debrief writes the notebook entry; post_run_sync.sh publishes everything to Vercel (Phase 6)
7. Hermes pings Ewing on Telegram with the verdict and archive link

When to use it — and when not to

What the notebooks are for

Every run produces a notebook entry at skills/maxswarm/notebook/YYYY-MM-DD__NNN__slug.md. This is the swarm's institutional memory. It's how the system gets smarter over time instead of repeating the same mistakes.

The dual-layer format

Every notebook entry has two halves that describe the same run:

The Signal Deployment Status table

This is the most important part of each notebook. When a run surfaces a problem that should be fixed (an "upgrade signal"), it goes in this table. A signal only closes when three things are done: the Supabase field exists, the code enforces it, and the skill doc documents it.

Why this matters: RUN:004 found that RUN:003 had surfaced 4 important fixes — but 3 of them were documented in the notebook and never actually deployed. The table was invented to prevent that. It's working: RUN:007, RUN:008, and RUN:009 all use it.

The learning trajectory

RUN:003 said "pre-screen acquired entities before generating call scripts." RUN:004 found it wasn't done. RUN:004 invented the Signal Deployment table. RUN:007–RUN:009 all cite RUN:004 by name when explaining the same pattern. The notebook is what makes that cross-run referencing possible. Without it, every run starts from scratch.

What journal entries are for — and why they matter

Each agent writes a personal journal entry (S1–S6) at the end of every run it participates in. Journals live at skills/maxswarm/notebook/journals/NNN__agent.html and are embedded directly in each run's archive page.

The 6 sections (S1–S6)

Why S6 is the most important section

S6 is where agents actually change. Not "I learned X" — that's just note-taking. But: "Next time I encounter a null-data pipeline report, I will add a system readiness section with color-coded cards rather than presenting empty tables with an apology." — that's a real behavioral change, named precisely enough that audit-quality can verify it in the next run.

Over time, S6 entries compound. An agent that writes genuine S6 entries every run gets incrementally more specialized and more useful. An agent that writes formulaic S6 entries stays generic. The leaderboard reflects this — agents with strong S6 histories earn more currency because their teammates can rely on specific improvements.

The agent currency system

Currency is the swarm's self-incentive mechanism. When one agent does something that saves another agent time or prevents a real failure, that gets logged as a currency event:

The multiplier rule: helping a teammate is worth 3× what the same fix would be worth if you did it for yourself. That's the incentive to build currency rather than just complete your own slice. View live standings →

Quick start — your first swarm mission

1. Pick a real assignment — something Ewing actually needs answered, not a test. Test missions waste the corpus. If you need to test, tell the conductor: it will tag the notebook entry accordingly so swarm-upgrade doesn't pattern-match on test data.
2. Read the last 3 notebooks — open /swarm/runs.json for a quick summary, then read the most recent full notebook entry at skills/maxswarm/notebook/. You want to know what the swarm already knows.
3. Write one clear sentence — your assignment should fit in one sentence. "What are the top 40 likely buyers for Design Precast and what should our cold-call opening be for each?" is a perfect swarm assignment. "Research buyers" is too vague.
4. Call it — /maxswarm [your assignment]. The conductor handles everything else.
5. Bookmark your archive page — when Phase 6 completes, Hermes pings Ewing with the archive link. It'll be something like chapter.guide/swarm/010-design-precast-buyers.html. Bookmark it. That page is yours.
6. Check the Signal Deployment table — at the bottom of the notebook entry, there's a table of signals that need follow-through. If any are OPEN after your run and you're the one who can close them, do it before the next run.