"Inside the AI Workflows of Every's Six Engineers" - Rhea Purohit

Why this is in the vault

This is a snapshot of how six engineers ran four AI products, a consulting arm and a 100k-reader newsletter in late 2025. It is useful less for the tool picks, which have since churned (the piece was updated 2026-09-17), than for the range of human-in-the-loop postures: from many parallel sessions to one session watched closely. Full text from the paid re-read.

The core argument

There is no single stack. Each engineer tuned the tools to their own temperament, and the shared pattern is plan → agent executes → deliberate review, with the human's attention as the scarce resource.

Yash Poojary (Sparkle) - experimenting at the edge.

Added a Mac Studio and runs Claude Code and Codex side by side on the same prompt. He calls Claude the "friendly developer" and Codex the "technical developer" (more literal, often right first try).
Figma MCP replaced screenshot-pasting, so the agent reads the real design system.
Keeps a "learnings doc": two lines after every push, stored in the cloud and fed back as rolling context.
Built AgentWatch to get pinged when a Claude Code session finishes.
Splits his day: mornings build (only Claude Code and Codex, no new tools), afternoons explore. "Easy to get derailed" is the reason for building in guardrails.

Kieran Klaassen (Cora) - orchestrating the loop.

Every feature starts as a Claude Code plan, scoped at three sizes: small (one-shot), medium (a few files plus a review step), large (manual typing, deep research, back-and-forth).
Context7 MCP grounds plans in current docs.
Plan goes to GitHub, then a "work" command turns it into tasks, then a review command runs (Claude plus Cursor/Charlie). This loops until he calls it shippable. This is the compound-engineering loop in practice.

Danny Aziz (Spiral) - turning complexity into milestones.

About 70% of his work is in Factory's Droid CLI: GPT-5 Codex for big builds, Anthropic models to refine.
In planning he asks the model about second- and third-order consequences (e.g., a DB access pattern that will slow the app) and turns them into milestones.
Warp for the terminal, Zed for reading plans. Hasn't opened Cursor "in months."

Naveen Naidu (Monologue) - process as source of truth.

"If it's not in Linear, it doesn't exist." Every ticket links back to its origin.
Two tracks. Small fixes: ticket context pasted into Codex Cloud. Big features: a local plan.md as the authoritative spec.
Uses Codex Cloud for exploratory draft PRs that are not meant to merge, to surface edge cases in parallel. The real build happens in Codex CLI, watched closely.
Three-step review: automated /review, then a manual before/after diff, then for bug fixes, Sentry error rates before vs. after the change.
Dictates prompts with Monologue.

Andrey Galko (engineering lead) - perfecting what works.

Sticks with what works. He moved from Cursor to Codex only because of pricing limits.
GPT-5-Codex closed the UI gap. Claude is still more creative, "sometimes too creative."

Nityesh Agarwal (Cora) - focusing on one thing.

MacBook Air, Claude Code Max, one terminal. Spends hours planning before any code.
Watches Claude "like a hawk," finger on Escape. Has shortened the leash, interrupting mid-run to ask for explanations. This means fewer hallucinations and keeps his own skills sharp.
GitHub is the team interface: humans leave line comments on Claude-authored PRs, and Claude fetches them and fixes.
Admits single-vendor risk: when Claude glitched for two days, nothing else matched.

Mapping against Ray Data Co

The constraint the piece circles is the same one Ray hits at Level 5: review bandwidth. Every engineer's workflow is designed around protecting human attention. Yash uses AgentWatch and a build/explore split, Nityesh runs one session, Naveen uses Sentry-verified review, Kieran sizes plans so only medium and large features get human review. None of them claims to review everything. Ray's fleet of subagents plus fresh-eyes critics is the automated version of Kieran's review command. The open question for Ray is Kieran's three-tier sizing: which outputs skip founder review entirely and which ones must reach him.

Naveen's Sentry before/after check is a behavior gate, not a reading gate. It is the same idea as the behavior-critic skill: verify that the fix reduced the error, not that the diff looks right. For the factory, the counterpart is comparing eval or error rates before and after a skill change.
Yash's learnings doc is a lightweight version of Ray's vault-plus-memory compounding, and the same idea as the Compound step in [[2026-02-09-every-compound-engineering-guide]].
Exploratory draft PRs that are never merged are a cheap way to explore options in parallel before committing to one. Ray could dispatch a few throwaway attempts before the real build.
Nityesh's single-vendor worry is relevant to Ray, which runs on Claude end to end. [[2026-07-25-multi-agent-claude-codex-grok-composition-patterns]] covers the hedge.

Why it matters for RDCO / The Denominator

Do: borrow Kieran's three-tier sizing as an explicit routing rule for Ray's outputs. Small means critic-only, no founder ping. Medium means critic plus a one-line HQ entry. Large means founder review. Write it down so review load is designed, not accidental.
Do (factory): add a before/after metric check (eval pass rate, or downstream error count) to the skill-change review, modeled on Naveen's Sentry step.
Write (The Denominator): "Six engineers, six leashes" is a usable angle. Successful agent deployments differ in tooling but share a deliberately chosen review posture. The denominator is attention, not model choice.

⚠️ Sponsorship

House promo throughout. Four of the six subjects are general managers of Every products (Sparkle, Cora, Spiral, Monologue), and the piece plugs AgentWatch, Monologue and the consulting arm, with a footer promoting Every's product bundle and Compound Engineering. The page also carries a Wealthfront cash-account disclaimer, which indicates a paid ad slot in the original email. Bias implication: tool endorsements double as product placement, and the tool picks are dated (late 2025). Weight the workflow patterns, not the vendor choices.

[[2026-02-09-every-compound-engineering-guide]] - Kieran's plan → work → review → compound loop, formalized
[[2026-01-28-every-stop-coding-start-planning]] - why planning comes first
[[2026-01-26-every-claude-code-shipping]] - Kieran's Claude Code workflow in depth
[[2026-02-02-every-codex-vs-claude-code]] - the Codex vs Claude comparison these engineers live in
[[2026-05-31-every-how-we-work-now]] - later snapshot of the same team
[[2025-08-14-every-claude-code-qa]] - Claude Code Camp Q&A with the same engineers
[[2026-07-25-multi-agent-claude-codex-grok-composition-patterns]] - composing multiple coding agents