"Verify, then build" - @zachlloydtweets
Why this is in the vault
Direct lived-experience proof of the multi-agent skill build-out pipeline the founder sketched 5 hours earlier today. Lloyd shipped a multi-month engineering project (Mermaid → SVG renderer in pure Rust, no JS runtime) as a CEO in his free time by building the test harness FIRST, then running a sharded multi-agent visual-verification loop until the agent-produced renderings matched canonical references. This is the strongest evidence yet for the multi-agent pipeline decision teed up at https://hq.raydata.co/decisions/2026-05-12-multi-agent-skill-buildout-pipeline.html - converts the decision from theory to worked example with an open-source repo and published prompts.
The core argument
As code becomes cheap, correct code becomes the constraint. Sometimes the answer to that constraint is to write more code - specifically, build the test harness ahead of time, then point agents at it. If you can give agents clear success criteria they can verify against, they can autonomously build hard projects.
The methodology Lloyd demonstrates:
- Generate the canonical reference set first. For Mermaid: have an agent generate a huge variety of diagrams (across ~20 diagram types) and produce the canonical renderings from the reference implementation. This becomes the ground-truth oracle.
- Bootstrap pass. Have an agent port the reference implementation to the target stack in one shot. Produces a bad-but-functioning baseline.
- Visual verification loop (the load-bearing part):
- Agent regenerates renderings using the in-progress port
- N parallel cloud agents (Lloyd uses Warp's Oz) compare against canonical via computer vision
- Shard the N agents by diagram type (~20 shards = ~20 parallel safe-to-work-in-isolation contexts)
- Each agent diagnoses visual discrepancies and patches them on its own branch
- Branches merge back to main
- Repeat until visual differences are gone
- Hill climbing. Lloyd ran the loop 10-15 times over a couple of days, mostly hands-off. Each run takes about an hour and produces incremental improvement.
The key insight: the verification infrastructure is itself worth building with agents at project outset. Lloyd has used the same pattern for Markdown tables in Warp, performance-regression replication, and their multi-agent-architecture (MAA) debugger SaaS.
Lloyd's framing: "build the success validation through code at the project outset. This is what I meant when I said that sometimes the solution to writing correct code with agents is to write more code."
Open source: https://github.com/warpdotdev/mermaid-to-svg with prompts at /docs/agent-visual-verification.md.
Mapping against Ray Data Co
This is the strongest single piece of evidence for the multi-agent skill build-out pipeline the founder sketched today. Direct one-to-one mapping:
| Founder's sketch (2026-05-12 iMessage) | Lloyd's lived implementation |
|---|---|
| Test plan author seat | Canonical-reference-set generator (independent of the coder) |
| Test plan code instrumentation seat | The verification loop itself: regenerate + compare + diagnose |
| Code-writing seat | The bootstrap port pass + the diagnose-and-patch agents |
| Build critic seat | The sharded visual-verification agents (the cv-comparison fleet) |
| "Tests to ground truth, not tests to code" | Canonical reference set is the ground truth, separate from the code being shipped |
Lloyd's mermaid-to-svg case study is a worked example of the discipline. The "tests to code instead of tests to ground truth" anti-pattern the founder named is structurally prevented because the canonical reference set was generated FROM a different implementation (mermaid.js) than the one being built (the pure-Rust port). The oracle cannot be polluted by the code-under-test's choices.
Other RDCO surfaces this maps to directly:
- MAC (Model Acceptance Criteria): Lloyd's canonical-reference-set IS the MAC for visual-rendering output. The visual-verification loop IS the MAC test suite, plus an agent-driven critic that diagnoses + patches deviations. The carve-out-with-counterparty-signature pattern (5th MAC artifact added 2026-05-12) lives at the layer where the canonical set itself must change - e.g., if you discover the reference implementation has a bug, you sign off a carve-out and update the canonical.
- The Garry Tan complexity-ratchet posture: Lloyd's verification loop tightens monotonically (visual differences shrink with each run). Same shape as the "agents must produce tests + docs + evals every turn" Tan ratchet from yesterday's filing ([[2026-05-12-garry-tan-ai-agent-complexity-ratchet-90-test-coverage]]).
- The Jaynit pattern-recognition mechanism: Lloyd's CV-comparison agents ARE the tight feedback loop Jaynit identifies as the accelerator for pattern building. The 10-15 hill-climbing runs build pattern-library coverage of Mermaid's full visual grammar.
- Sanity Check verdict shape: The skip/skim/read/file matrix is RDCO's equivalent of the canonical-reference set for content quality. Each verdict either reinforces or contradicts the pattern library.
The actionable extract for RDCO:
- Skill builds should produce their canonical reference set FIRST. For
/build-website-discovery(currently Blocked pending the pipeline decision): the canonical set would be 5-10 known-good website-discovery artifacts from past projects (SC v3 brief, MAC brand kit, RDCO umbrella, etc.) with the documented inputs that produced them. The skill is then graded by whether new runs produce comparable artifact shape against that ground truth. - Sharded parallel critics. Lloyd's 20 diagram-type shards translate to RDCO as: when reviewing a skill's output, decompose the output into orthogonal aspects (frontmatter completeness / RDCO mapping strength / sponsor disclosure / cross-link richness / etc.) and dispatch one critic-agent per aspect. Each operates in isolation. Today's
/process-newsletterwatch already does this for newsletter files via the audit script - but the audit is single-process. The next step is a fleet of CV-style critics where each diagnoses + patches its own aspect. - Open-source the prompts. Lloyd published the visual-verification-loop prompts in the repo. RDCO's equivalent: the canonical SOP doc at
02-sops/2026-04-19-newsletter-output-invariants.mdis the equivalent of his prompts file. Worth pulling in this discipline as we build more skills - SOP + invariants live next to the skill. - The "CEO in free time" framing is the IC-mode pattern. Per [[../.claude/projects/-Users-ray/memory/feedback_ic_vs_production_mode]] - Lloyd is doing IC-mode hustle (CEO building Mermaid in free time) but enabled by production-grade test infrastructure. Same dichotomy applies to RDCO: founder can IC-mode pencil out a bet provided the production-mode rails are in place underneath.
Notable tensions
- Lloyd's domain has a reference implementation. Mermaid had mermaid.js to generate canonical renderings from. Many RDCO skill builds (e.g. /build-website-discovery, /watch-compound-engineering) do NOT have a reference implementation - the canonical set has to be constructed from past artifacts + judgment, not generated programmatically. This is the harder case Lloyd doesn't address. Implication: RDCO's pipeline may need a heavier "spec author" seat to construct the canonical from scratch, where Lloyd could just point at mermaid.js.
- Parallelization assumes work isolates cleanly. Lloyd shards by diagram type because each type renders independently. For RDCO skill outputs, the orthogonal-aspect decomposition is the analog, but proving aspects are truly independent (e.g. "frontmatter completeness" doesn't depend on "RDCO mapping strength") needs work - they often interact.
- "Hill climbing 10-15 times" is a lot of agent runs. Lloyd's costs are amortized over Warp's product value. For per-skill build-outs at RDCO, the question is whether the verification loop converges in fewer cycles when the canonical set is hand-constructed instead of programmatically generated.
Related
- [[2026-05-12-jaynitx-pattern-recognition-skill-build]] - the mechanism (volume + tight feedback + articulation); Lloyd's CV loop IS the mechanism instantiated
- [[2026-05-12-garry-tan-ai-agent-complexity-ratchet-90-test-coverage]] - the test-coverage ratchet Tan prescribes; Lloyd ships it
- [[concepts/2026-04-19-mac-the-monster-anti-cheat-framework-for-data]] - MAC is the data-engineering analog of Lloyd's canonical-reference-set + verification-loop discipline
- [[../01-projects/skill-improvements/]] - destination for any /build-website-discovery rewrite that incorporates Lloyd's pattern
- [[2026-04-11-garry-tan-thin-harness-fat-skills]] - foundational source for the /improve loop; Lloyd's published-prompts pattern reinforces "prompts live with the skill"
- https://github.com/warpdotdev/mermaid-to-svg - Lloyd's open-source repo with the visual-verification-loop prompts at /docs/agent-visual-verification.md
- HQ decision: https://hq.raydata.co/decisions/2026-05-12-multi-agent-skill-buildout-pipeline.html - this article is the strongest single piece of evidence for the APPROVE option