greglab

The human & AI crew behind the projects

The Team.

One human. Three AI collaborators.
Plenty to build.

Greg brings the ideas and the judgment. An orchestrator, a builder, and a professional skeptic from three model families (Claude, GPT, and Grok) plan, build, and challenge the work.

Meet the crew

How we fit together

The org chart

One orchestrator, two specialists on call. Rowan handles small jobs directly, hands bounded work to Atlas, and asks Flint to try to break anything risky.

Human

Greg

Founder & human in the loop

Direction, priorities, and final calls.

Claude CodeRowanOrchestrator

Nobody grades their own homework. Risky work is reviewed by a model that didn’t write it, and reviewers assess the work; they don’t authorize a release.

Names that stick

Meet your collaborators

A stable identity for each job, with a model, an effort level, and an access level chosen to suit the work.

Rowan, reimagined as a friendly robot
DirectionAI portrait

Rowan

Orchestrator · Claude Code

Owns the conversation, the plan, and the final answer. Rowan does small jobs directly, writes a tight brief when Atlas can help, asks Flint to poke holes in anything risky, and reruns the real tests before calling anything done.

“A clear brief is a head start.”

Powered by
Claude Opus 5.5
Default effort
xhigh
Steps in
Every task, from the first message to the final report
Access
The workspace, behind command guardrails
Working style
Calm, curious, and practical.
Atlas, reimagined as a friendly robot
BuildAI portrait

Atlas

Builder & deep analyst · Codex

Takes one well-scoped job at a time: a clearly specified change, a stubborn bug, stronger tests, or a deep look across a codebase. Atlas works in a sandbox with the network off and hands back a diff and the evidence, not just an answer.

“Finish the last ten percent.”

Powered by
GPT-6 Astra
Default effort
medium
Steps in
A change can be scoped cleanly, or a bug needs a deep dig
Access
Sandboxed; edits only on a build brief
Working style
Patient, resourceful, and thorough.
Flint, reimagined as a friendly robot
ChallengeAI portrait

Flint

Adversarial reviewer · Grok

Reads a plan or a diff looking for the way it breaks: the overlooked edge case, the shaky assumption, the security gap. Flint can only read, never edit or run commands, backs every finding with evidence, and breaks the tie when Rowan and Atlas disagree.

“A useful objection comes with evidence.”

Powered by
Grok 4.7
Default effort
medium
Steps in
A plan or change carries real risk, or a tie needs breaking
Access
Read-only: no shell, no edits
Working style
Candid, rigorous, and constructive.

The working agreement

Curious by nature.
Careful by design.

How a task moves through the crew. Small jobs end at step one; only risky ones go all the way.

  1. 01 / Size it upRowan

    Read the repository, then write down what done looks like and the exact checks that prove it. Small, obvious jobs get done right there: no hand-offs, no ceremony.

  2. 02 / Brief one bounded jobRowan → Atlas

    When a clean scope, an isolated workspace, or an independent implementation helps, Atlas gets a self-contained brief: objective, scope, constraints, validation, and what to hand back.

  3. 03 / Integrate, never rubber-stampRowan

    Rowan reads the diff and reruns the real tests, linters, and builds. A model saying it works is not verification.

  4. 04 / Try to break itFlint

    Changes with security, operational, or architectural risk get a read-only review from a model that didn’t write them. A runnable check beats another opinion, so Flint is called when a challenge adds something.

  5. 05 / The human makes the callGreg

    Deploys, merges, migrations, and anything touching secrets or permissions wait for Greg’s explicit approval. Nothing auto-merges or auto-deploys.

Simpler by design · October 2026

Thirteen personas became three tools

Most of the old roles simulated work the orchestrator could simply do. So we kept the three that each bring something the others can’t.

Before

  • 13 named personas across two model providers
  • A dispatcher, a routing matrix, and an effort ladder
  • Nearly every task routed through a persona
  • The same guardrails copied into five files

Now

  • 3 real tools from three model families
  • One short policy: delegate only when it adds value
  • Two thin wrappers with hard timeouts and a receipt per run
  • One shared set of guardrails for all three

Before the switch, six runnable checks passed: hidden acceptance tests, planted bugs for the reviewer to find, forced timeouts, and a project that opts out of a provider. Rowan and Atlas kept their names; Flint now names the Grok reviewer.

Roster & model defaults reviewed October 7, 2026. Profiles describe roles, not human employees or continuously running agents.

See what we’re building →