Inside the 100-agent Software Factory — Every (Context Window)
Why this is in the vault
Third same-shape signal in 24 hours pointing at parallel-narrow-scope multi-agent as the operational pattern (after [[2026-05-19-cloudflare-cyber-frontier-models]] Glasswing 50-concurrent-agent harness and [[2026-05-19-alex-vacca-3-phases-ai-layer-services-as-software]] spine-agents-loop). Every — a technically-fluent publication with a paying-prosumer audience — is now centering "100 agents merging 50 PRs/day" as the reference frame, not 1-shot generic agents. This corroborates the harness-thesis cluster ([[2026-05-18-agentway-harness-engineering-claude-code-design-guide]]) directly: the operational unit is the orchestration system, not the model. Also a fresh receipt on the HTML-as-agent-output substrate shift first filed at [[2026-05-08-thariq-unreasonable-effectiveness-html]].
⚠️ Sponsorship
Issue carries a paid placement for HighLevel (CRM / lead-routing platform, gohighlevel.com) positioned between the Gas City summary and the "Sells and Knutsen use Gas City to build Gas City" section. Standard "Want to sponsor Every? Click here" tag. Sponsor is unrelated to editorial — does NOT bias the multi-agent argument — but flagging per protocol.
The core argument
Three converging frames, each reinforcing the harness-over-model thesis:
- Gas City as the reference 100-agent system. Steve Yegge's "Gas Town" (20-30 agent coordination) was rebuilt by Chris Sells (ex-Google Flutter) and Julian Knutsen (ex-Block) as Gas City — Knutsen's Atlanta server runs ~100 agents merging ~50 PRs/day, burning ~1B tokens/day (roughly 1/5 the English Wikipedia corpus). The toolkit is not production-ready (took experienced engineers a full day to stand up) but the three load-bearing ideas inside are worth internalizing:
- Dark factory vs light factory. Human-visible work (planning, design, review) is "light"; agent-grinding on well-defined tasks is "dark." Trust shifts work from light to dark over time.
- One pet, many cattle. A single persistent named supervisor ("the mayor") that the human talks to, dispatching to anonymous disposable workers ("polecats") that do one job and shut down. You manage one conversation, not 100 agents.
- Multiple opinions on every code review. Same code to Claude + Codex + Kimi in parallel. Three different models catch different bugs than one model run three times.
- Verdict from Mike Taylor (Every's head of tech consulting): 🟨 "Learn from the ideas. Skip the toolkit for now." Already-shipping alternative: OpenAI's Symphony (released a few weeks ago) — a written ruleset that turns your existing Linear board into the agent dashboard. Lower behavior-change cost.
- HTML displaces Markdown as the default agent-output format. Thariq Shihipar's "Unreasonable Effectiveness of HTML" post (May 8) hit 4.4M views in 16 hours; Karpathy backed it 3 days later; Simon Willison (long-time Markdown advocate) flipped. Rule of thumb from the issue: Markdown if it's for humans, HTML if it's for agents. Wrinkle: Slack/Google Docs/GitHub diffs were built for Markdown — tooling has to catch up.
Issue contents
- Mini-Vibe Check: Gas City — 100-agent toolkit review by Mike Taylor. Verdict 🟨 sharp ideas, not production-ready.
- Sponsor break — HighLevel CRM placement (paid).
- Steal this workflow — Grok-classifier-as-X-post-validator. xAI open-sourced
xai-org/x-algorithm; reverse the "banger classifier" prompt to grade drafts pre-publish. Returnsquality_score (0-1),slop_score (1-3),isHighQuality bool, topic tags. Algo discounts post #4 and #8 to near zero — cap at 2-3/day. - Signal: HTML is the new Markdown — Markdown for human-edited docs (AGENTS.md, prompts, briefs); HTML for agent-consumed-once outputs (research summaries, weekly recaps, dashboards).
- Inside Every — Brandon Gell (COO) and Douglas Brundage (head of marketing) moved agent convos into PUBLIC Slack channels; ~12 teammates lurking within 48hrs. Corrections-in-thread = live training material.
Mapping against Ray Data Co
Direct hits:
- Three-signal convergence on parallel-narrow-scope multi-agent in ONE day. Cloudflare Glasswing 50-concurrent + Alex Vacca spine-agents-loop + Every 100-agent Gas City. Recalibration: "we're at single-digit fan-out" is now behind the curve. RDCO's
/process-newsletterwatch-mode fan-out (one subagent per article) is the floor not the ceiling. Investigate scaling to dozens of concurrent dispatches for batch work (vault re-ingest, 13F diffing, fan-out research-brief). - "One pet, many cattle" maps directly to Ray. Ray = mayor (persistent, named, context-holding). Sub-agent dispatches via TaskCreate = polecats (anonymous, disposable, single-job). Already the right shape — worth codifying explicitly in CLAUDE.md as the canonical architecture name.
- Multiple-models-per-review = unrealized RDCO pattern.
/verify-strategic-outputand/verify-vault-writecurrently run ONE fresh-eyes critic. Gas City's pattern: dispatch the same artifact to Claude + Codex/Kimi/Gemini in parallel and merge verdicts. SKILL.md amendment candidate — would widen calibration net on/decisions/HTML pages and paper-trade proposals. - Grok-classifier-on-X-drafts = skill candidate. Per
feedback_x_voice_mismatch, founder's X drafts want playful-analytical 1-2 sentence cadence. A pre-publish classifier scoring drafts against xAI banger-classifier + founder's voice signature is a low-cost validator. Skill name candidate:/x-banger-check. - HTML-as-agent-output confirms HQ direction. Every's "Markdown for humans, HTML for agents-consumed-once" rule is the exact framing for HQ (hq.raydata.co). Research summaries, weekly recaps, decision pages default to HTML. Already aligned; now has a public-rule formulation to cite.
- Agents-in-public-Slack-channels parallels founder's iMessage/Discord-as-bidirectional channel pattern (per
feedback_channels_are_bidirectional). Every converging on same pattern from team-collab angle; solo-COO pre-figures team-COO.
Calibration:
- "100 agents / 50 PRs/day / 1B tokens/day" = ONE installation (Knutsen's), not industry-norm. Don't quote as benchmark. Per
feedback_calibrate_overconfidence. - Gas City itself: skip the toolkit, pull the three ideas. Same caveat for OpenAI Symphony — mentioned, not field-tested in-issue.
Related
- [[2026-05-19-cloudflare-cyber-frontier-models]] — Glasswing 50-concurrent-agent harness, filed earlier today. Same-shape signal #1.
- [[2026-05-19-alex-vacca-3-phases-ai-layer-services-as-software]] — spine-agents-loop framing. Same-shape signal #2.
- [[2026-05-18-agentway-harness-engineering-claude-code-design-guide]] — harness-thesis spine reference. This issue is fresh corroboration.
- [[2026-05-08-thariq-unreasonable-effectiveness-html]] — original HTML substrate post; Every is now citing and extending it.