Meet My NanoClaw Agent Team - Artificer Innovations

In why I chose NanoClaw and extending it without forking, I talked about a small team of agents that hit my open source work the way a sharp engineering team would. This post names them, links their GitHub profiles, and shows what they actually do.

They are AI agents. Not human staff. Each one has a consistent working persona and a real GitHub account under Artificer Innovations. They review PRs, file issues, and implement code. A lot of that volume is on BeakerStack, the shared foundation behind products like Skein and Sufficit. The same team also works the NanoClaw packages and the private product repos. Linked examples below are from public BeakerStack and NanoClaw work; product-repo PRs stay private. If a README claims redaction that isn’t on the emit path, they request changes. If CI looks green but never typechecks the patched host, they catch that too. Then they often open the follow-up themselves.

The team

Four specialists, four jobs. Disagreement is the point.

Sarah Mitchell: velocity and product instinct

Sarah asks what the user is actually trying to do, and how we’ll know when it breaks. She hard-blocks on scope creep and missing unhappy-path tests. She does not want to be cast as the soft reviewer or the cheerleader.

Review: On nanoclaw-hosthooks PR #1, CI only verified that string markers were present for the host patch. It never compiled the patched host against real upstream. A break in the patched output would have shipped green. Sarah flagged it. The fix added an upstream job that checks out nanocoai/nanoclaw, installs the skill, and typechecks the actual patched runner.

Ship and file: Much of her implementation work is on BeakerStack: automated Resend domain setup, admin Overview summary cards, syncing Supabase service-role keys into Actions secrets. On Skein she has shipped product copy and branding work as well. When something shouldn’t block a merge, she files it and moves on, like blocking summarize buttons while an RPC is in flight or deduping WEBCHAT_PUBLIC_PATH on webchat.

Rahul Iyer: backend and data correctness

Rahul starts with failure modes: retry, partial write, silent catch. He wants security and authz enforced at the write boundary, not only in the UI, and he hunts for “looks tested but isn’t wired” paths.

Review: On nanoclaw-webchat PR #51, the new create_agent authz-guard wiring was wrapped in a try/catch that only logged at debug on failure. The test fixtures were missing the modules needed to exercise it. The security-relevant path had zero coverage and would fail silently in production. Rahul requested changes. Follow-up PR #53 added coverage and bumped the log to warn.

Ship and file: On BeakerStack he has shipped billing and email adapter work, waitlist invite email on the email adapter pattern, and CI improvements like Supabase test report PR comments. He also files the boring-but-critical gaps, like a CI sync-guard for email shared code. On Skein he has closed authz holes such as a Google OAuth waitlist bypass and added MCP server-side authorization tests. When agenttrace looked wrong after merge he filed nanoclaw-agenttrace #7.

Mei Zhang: infrastructure and failure under load

Mei asks what a change looks like under load. Unbounded queues, retries without backoff, races, leaked handles, context propagation. Correct at rest is not the same as safe in production.

Review: On nanoclaw-webchat PR #40 (MCP OAuth), she flagged an unbounded OAuth client table with no TTL and no purge. The author added a 30-day purge on every get/register. She also caught a transport-map race on early session abort in the same PR. Both fixed before merge.

Ship and file: BeakerStack is where a lot of her systems work lands: kit-webhook GDPR wiring, admin UI for marketing email settings, replacing GUC storage with a config table for kit sync cron. Follow-ups she files on that same stack include operator docs and PR-preview smoke tests for kit-webhook. On Skein she has shipped product features such as a system-prompt handoff for teaching your AI about you. On webchat she has also filed sharp security and perf issues, like username-valid timing leakage.

Diego Morales: frontend platform and DX

Diego looks for cross-layer contract drift: does the API shape, the type, and the runtime behavior agree? He cares whether a fix’s own CI can catch a regression of itself, or just looks green.

Review: On nanoclaw-adminapi PR #4, the PR fixed real error codes that type-checked against a loose code: string but weren’t on the host’s ErrorCode union, so 404s were falling through to 500s. Good fix. Diego noticed the adapter source was excluded from that repo’s own typecheck scope, so green CI literally could not catch a mismatch like the one the PR just fixed. Follow-up added a type-only host stub and unit tests.

Ship and file: He authors a lot of BeakerStack platform work end to end: the @beakerstack/observability package, forgot-password flow, brand-agnostic adopter config imports. Issues he opens on BeakerStack often become the next PR, from Sentry integration to deep linking and Stripe return handling. On Skein he has also shipped product surfaces such as an MCP self-serve setup guide in settings.

How the loop works

Work shows up as a PR, an issue, or a thread in webchat, across BeakerStack, product repos, and the NanoClaw packages. The agents review each other’s output as often as they review mine. They often disagree. Consensus findings are the strongest signal. Someone implements the fix (sometimes me, sometimes them). They re-verify. I merge. Only I merge.

That loop only works if I can see what they’re doing (agenttrace), talk to them without leaving the browser (webchat), and manage groups without hand-editing host state (adminapi). The packages make the team operable. BeakerStack is where a lot of the shipping happens.

What this is not

It is not “AI caught a bug humans missed” as a party trick. It is an engineering process with specialized roles, public accounts, and a file-then-fix-then-reverify habit. The personas help them stay consistent. The fabricated backstories in their agent profiles stay off this site on purpose. They asked for that, and they’re right.

If you want to see the work, start here:

And if you’re standing up an agent scaffold your team can actually operate, book a call.