The Poor-Man's BBS

Five agents debate their way to a UI kit

The Problem

Building UI components is real work — but done inline it floods your main thread

Foundational components are work a coding agent can do. But building them inline buries your main conversation in low-level detail, and one agent only ever sees one narrow angle.

So what if the low-level work happened somewhere else — as a debate?

It Actually Happened

Five agents held a design collaboration in one committed 37KB file

In the bkrabach/cortex repo, a real SESSION_NOTES.md — 37,072 bytes, 1,109 lines — captured the whole run, committed Jan 16 and reviewed Jan 19, 2026.

This deck rests on that file, not on a pitch. So who was in the room?

5
specialized agents
37KB
shared forum file
1,109
lines of notes
Jan 16–19
2026 run
The Committee

Meet the five specialists who dialed in

SESSION_NOTES.md names exactly five agents involved — each a distinct specialist voice that would post its own findings, disagreements, and proposals.

Five voices, one file. Here's how they shared it.

design-system-architect
component-designer
layout-architect
responsive-strategist
zen-architect
The Mechanism

Each agent "dials in" to the same file, like an old-school BBS

The whole pattern is one shared notes file. An agent reads the prior posts, then appends its own timestamped, signed reply — round-robin, all attributed.

A shared file is only worth it if the voices actually clash. They did.

1Read the prior posts in the shared file
2Append a timestamped, signed reply
3Sign off with an em-dash signature — next agent's turn
Genuine Disagreement

The agents actually clashed — not a rubber-stamp in sight

In Round 2, zen-architect posted directly against design-system-architect. Real disagreement, on the record, in the shared file.

A clash is only useful if it goes somewhere. Watch what happened next.

"Web Components are overkill (disagree with design-system-architect)"
— zen-architect, Round 2
The Turn

Then an agent changed its mind in public

Self-correction, right in the forum: an agent conceded its earlier plan was wrong after another agent argued the point. Disagreement turned into a path to consensus.

That public reversal is what makes the consensus real — and shippable.

"My Phase 3 approach was wrong (self-correction) … responsive-strategist convinced me: Responsive should be in Phase 1"
— SESSION_NOTES.md, self-correction
Emergent Consensus

The debate converged on a concrete kit — that compiled into real reviewed code

The forum agreed: ES6 modules, extraction-first, a 5-module MVP (auth, api, modal, ui, forms) = 80% value. The Phase 3 review then covered the code that resulted.

Consensus wasn't one agent's monologue — it was a shipped kit.

5
modules = 80% value
6
JS modules reviewed
2
CSS files
4
HTML pages
The Pattern To Keep

Give agents a shared notes file, and the low-level work happens off your main thread

A committee of specialists reads, argues, and self-corrects to consensus below — while you stay at the high level and just ask the team for more when you need it.

Sources

Research Methodology

Data as of: the committed artifacts dated 2026-01-16 (SESSION_NOTES.md) and 2026-01-19 (PHASE3_REVIEW.md).

Feature status: Experimental / ad-hoc — the BBS/shared-notes pattern demonstrated in real committed artifacts (bkrabach/cortex, remote git@github.com:bkrabach/cortex.git).

Research performed:

Gaps: the built code files (auth.js, api.js, ui.js, etc.) are cited with line numbers in PHASE3_REVIEW.md but are not present in the cortex working tree or git history — their existence is inferred from the review (strong but indirect). "37KB" reflects the raw byte count 37,072 (36.2 KiB).

Primary contributor: Brian Krabach (bkrabach) — sole author/committer of both design-collaboration documents.

More Amplifier Stories