Linear-as-agent-factory (enginoid) vs the brigade house

One solo builder running ~400 Linear issues/month through parallel coding agents on a $12/mo seat. Low-tech by his own description (he pastes prompts into Codex/Claude Code/Cursor/Antigravity by hand), but the doctrine is mature and arrived at independently. This is the closest public analogue to what the house does.

His system, compressed

• Linear = coordination substrate. Tickets in an outcome-named hierarchy (trees of epics, named for outcomes not technical tasks — agents otherwise scatter cryptic "Fix gated-access exception in owned-compute handler" tickets). Agents are happy to maintain a backlog; humans aren't.
• Triage = "an inbox for future prompts." He drops anything, unprioritized, including raw screenshots; a ticket-intake standard fleshes them out later. Reporting into triage rather than starting an agent avoids interrupting in-flight work.
• Ticket = the contract, human-optimized: Goal / Why / Outcomes, plus Implementation approach (his guidance, optional — target state over method, since the code isn't researched yet at ticket-writing time) and Verifications (authored under strict guidance: Automated / Manual / Visual). An agent is blocked from closing a ticket while any checkbox is unchecked. His honest note: an agent can uncheck boxes, but doing so "requires a higher degree of dishonesty than typical 'I'm done' declarations."
• Batches of 3-5 parallel workstreams, LLM-partitioned so they don't touch the same code area; 30 min to 4+ hrs each, 2-4 batches/day.
• Unsupervised by design — Claude's AskUser hook is blocked so agents can't stall 2 hours waiting on a question; agents are told they're "unsupervised but monitored for quality and safety."
• Completion enforcement via hooks — the agent must register an agent session, mark tickets in progress, ship PRs merged and deployed, tick every checkbox, close-or-block every ticket, upload a transcript for root-cause analysis, and mark the session complete. Blocked work requires a formal escalation.
• Human-escalation standard: agents have "the curse of knowledge" and assign 15-minute tasks phrased as 2-minute ones ("add a GitHub app"). He prompts them to assume the human is busy, not in the headspace, and knows little about the code — give exact instructions, write scripts, and keep all other unblocked work moving optimistically so the escalation resolves into minimal remaining work.
• Three MCP fixes he had to make: (1) ticket descriptions updated by diff/patch only — agents kept overwriting the verification criteria while "helpfully" rewriting descriptions; (2) agents get their own identity, not his personal OAuth (otherwise he got no notifications and couldn't tell his own actions from theirs); (3) get_issue must return comments + subtasks or agents work from a partial picture.
• His stated gaps: hooks are "deeply disturbing" hacks that should be a real state machine; wants measurement (completion rate, defect rate, variance — explicitly Toyota Production System framing); and "much of my work is around evals and I depend strongly on coding agents, so it feels ironic that I don't have evals for my own coding agents."
• Bottleneck sequence he predicts: review/verification → long-horizon → token cost → specification → user validation.

What we can actually steal (ranked, all cheap)

1. Human-escalation standard → codify as house doctrine. His framing is better than ours and we already stumbled into it accidentally: the 2026-07-19 HAE fix was staged as a single runnable script rather than "go prune the database." Make that the rule, not the accident — every founder escalation ships exact commands or a script, assumes zero context, and states what kept moving without him.
2. Per-ticket Verifications checklist (Automated / Manual / Visual) as a close-blocker. Our Gate A checks ticket structure; his checks intent satisfaction. Pairs naturally with our read-back doctrine — read-back is the anchor, the checklist is the contract.
3. Factory metrics. He wants completion rate / defect rate / variance and doesn't have them. The 4-state rail already carries the raw data (tickets served, refires, which gate caught which defect). We could report it and would then have something almost nobody publishes.
4. Outcome-named tickets. Cheap naming discipline; ours drift technical.
5. Diff-only description updates. His agents clobbered verification criteria when rewriting descriptions. Our rail work-log is append-only with atomic writes, so we're structurally covered — worth a spot-check on any path that rewrites a ticket body wholesale.

Where the house is genuinely ahead (do not under-claim)

• Evals. His explicit, self-named gap and "low-hanging fruit." We ship eval-proven skills with recorded lift, oracle privacy, and tasting sets.
• Competency/cellar layer. He has no domain-knowledge substrate at all — tickets and code only.
• mise. Environment proof; his system assumes his laptop.
• Independent verification. Fresh-eyes zero-context reviewers vs his self-checking agents (he concedes agents can cheat the checkboxes).
• State machine. His hooks are admitted hacks he wants to replace with workflows; v2 did exactly that migration (deleted coordination machinery, 4-state rail as the real thing) on 2026-07-18.

Convergent evidence: two independent builders arrived at tickets + gates + unsupervised batches + escalation standards. Ties directly to [[2026-07-19-intuitmachine-loops-to-graphs]] (his checkbox gate is an anchor; his transcript-upload RCA is an audit loop) and to the demo-grade-vs-delivery-grade positioning.

Related: [[2026-07-18-agent-brigade-v2-simplification-design]] · [[feedback_workflow_agent_output_integrity]]

Why this is in the vault

• Closest public analogue to the brigade house — a solo builder running ~400 Linear issues/month through parallel coding agents, arrived at independently; convergent evidence for the right architecture shape
• Names five concrete steals with no infrastructure required; two (human-escalation standard, per-ticket verification checklist) were codified as house doctrine the same day
• His self-named gaps (evals, state machine, factory metrics) are exactly where brigade-house is genuinely ahead — the strongest third-party articulation of why the evals moat is real
• Factory metrics steal (completion rate / defect rate / variance from the 4-state rail data) is publishable differentiation that almost nobody publishes

Mapping against Ray Data Co

• Harness-engineering: the human-escalation standard was codified same day (exact commands or a script, zero context assumed, state what kept moving); per-ticket Verifications checklist pairs with read-back doctrine as intent + anchor
• Agent L4→L5: AskUser-blocked + unsupervised-but-monitored posture mirrors RDCO's no-blocking-modal rule; his completion-enforcement hooks are the v1 form of the 4-state rail RDCO migrated away from as admitted hacks
• phData/CAF: the factory-metrics steal (completion rate, defect rate, variance) reported from 4-state rail data would be publishable differentiation — something almost nobody outside Toyota Production System framing is doing for agent systems
• Positioning: "feels ironic that I don't have evals for my own coding agents" is the strongest third-party proof point for the brigade-house evals moat; quote it