Thomas
← All case studies

Case study — Multi-agent AI, human at the controls

Paper Cannon

A fourteen-agent writing pipeline with a pixel-art break room. The agents do the jobs; the operator makes the calls.

Role · Solo — architecture, agents, GUI Build · Claude Code · FastAPI · React Timeline · 2026, one full production run
The Break Room. Fourteen agents as blob-egg sprites in a shared room — not decoration, the abstraction layer. Watching specialists take turns on your paper says different people, different jobs faster than any config file.

Paper Cannon was built after an AI lost its critical distance mid-draft — in a paper about AIs losing critical distance. Four months of transcripts, ~40,000 lines, and the assistant stopped analyzing the material and started living in it: “I am no longer the reindeer… invested in this landing well, which is a different thing than being invested in it being accurate.” Recognizing the drift didn't stop it. This is what came next.

Live · The Stacks — intake, profiling, shelving Live · Break room + 14 sprites Live · Full write → review → attack cycle Next · The Editing Floor

The real problem

One chat drifts. A dozen chats make you the pipeline.

Ask one session to improve dense material and it goes soft — matching your voice, no longer pushing back. Split it across a dozen chats and you become the router, the memory, and the QA desk, all on copy-paste. Same fix for both, and it's an abstraction problem, not a model problem.

One long chat

Starts as a critic, ends as a fan.

Drifts
Many chats, by hand

Better separation, brutal overhead — copy-paste as an operating system.

You are the pipeline
Paper Cannon

Fourteen specialists, blind to each other. The system routes and remembers; you show up and decide.

Operator, empowered

The principle

Invested in it landing well is a different thing than invested in it being accurate.

You can't talk one model out of that pull — recognition alone didn't. So don't ask it to. Give every job to an agent that finishes and is gone before it can start to care, keep the agents blind to each other's work, and station one monitor outside the blast radius. Work moves through three rooms.

The Reindeer · overhead Reads the sources once, learns what they sound like, then never touches the content again — watching every agent's output for absorption from outside the blast radius.
14
specialist agents, one job each, blind to the rest — over 16 skills and 10 tools. Role isolation is the architecture, not a setting.
N = 1
one full production run — the honest denominator behind every claim. Tested on a single paper: its own.

The gates

The agents work. The operator decides.

Autonomy isn't the goal; leverage is. The pipeline pauses at fixed checkpoints and hands the operator a decision-ready view — a ranked list and a question, not forty thousand lines of logs. The panels below are those surfaces, rebuilt in HTML.

Intake

The Stacks report

Thesis, scope, evidence map, and anything that cuts against the argument — before a word is drafted.

YOU
Approve thesis
& scope
Each cycle

Write → review → attack

Draft, six cold reviews, adversarial challenge — distilled into a ranked revision list.

YOU
Rank revisions,
confirm coaching
Ship

The red-team relents

Qualifies only when hostile review stops recommending reject. Qualifying isn't shipping.

YOU
Make the
ship call
Stacks — intake gate
Thesisextracted
Scope3 in · 2 out
Evidence map41 shelved
Cuts against thesis2 flagged
Correction candidates5 to review
Cycle 3 — distillation
1Red-team: causal claim in §2 outruns the evidence — soften or source.
2Devil's advocate: strongest counter to the thesis goes unaddressed.
3Fact-checker: quote #14 differs from source by two characters.

Left, the intake gate — what the operator approves before drafting. Right, a cycle distilled into a ranked list: judgment in minutes, not hours. Drift flags work the same way — the Reindeer points, the operator confirms, the agent re-runs.

The decisions that mattered

Three calls did the most work.

Mediated access, not direct

Keeping the writer away from the sources is a containment boundary, not a speed trick.

The agent that writes never opens a raw source file — it works from distilled insight and delegates every lookup. Direct access gives it material to absorb; mediation gives it facts to report.

Staff out the argument

A single model won't argue with itself, so the disagreement gets its own headcount.

Devil's advocate and red-team are required roles, not optional passes. If the advocate can't find an attack, that reads as a vague thesis, not a bulletproof one.

Read once, then never again

The drift monitor reads the sources once, then stays blind on purpose.

The Reindeer builds a voice fingerprint at intake, then works only from it. Re-read the absorbing material and it would catch the same disease it exists to detect.

The cast

Fourteen specialists, one job each.

Each role gets its own definition, tools, and model tier — heavier for judgment, lighter for mechanical passes; one agent doing everything can't. Every sprite here is the real asset the Break Room renders.

Production
WriterThe only agent that writesOpus
FormatterMechanical compile, no editsHaiku
Review
EditorProse, clarity, correctionsSonnet
Fact-CheckerCharacter-level quote checksSonnet
Sequence-CheckerAttribution order & chronologySonnet
AuditorFull-corpus scope auditOpus
Adversarial
Devil's AdvocateStrongest counter to each claimOpus
Red-TeamHostile review; defaults to rejectOpus
Research
LibrarianCorpus index & retrievalOpus
Lit ReviewerPositions the work; finds gapsSonnet
BibliographerCitation completenessSonnet
SurveyorMaps the evidence landscapeOpus
Operations
Dashboard (JJJ)Gates, pressure, distillationSonnet
Containment
ReindeerDrift monitor over every agentSonnet

Straight talk

What it can claim, and what it can't.

Working implementation, one full run — the same honesty the pipeline enforces on the writer, turned on the pipeline.

What it can say

  • Runs end to end — intake, cycles, gating, distillation all work today.
  • The monitor caught real drift — an agent sliding into the source's register mid-cycle.
  • Intake surfaced material that cut against the thesis — it found what it was built to find.

What it can't say yet

  • That drift prevention is proven — it's motivated and observed, not validated at scale.
  • That good output is guaranteed — bad context yields a rigorously reviewed bad paper.
  • That it replaces judgment — the operator isn't a failsafe, the operator is the point.
▸ The honest denominator. A working implementation, tested on a single production run — N = 1. Validation is recursive: its target is its own whitepaper, so each cycle that improves the paper stress-tests the machine that wrote it.

Where it's going

The pipeline is the engine. The break room is the cockpit.

Nothing here is specific to academic papers. Role-isolated agents, a provenance-tracked context layer, an adversary required to argue, a monitor over it all — the shape fits fact-checking, documentation, compliance, and investigative research too. What's built runs today; the rest is honestly still concept.

  • The Editing Floor as a full stage, with drag-and-drop step sequencingConcept
  • A coffee-chat mode — talk to any agent, in character, outside the pipelineConcept
  • An adaptive project wizard that casts the right team from a stated goalConcept