Back to the LabLab architecture

How the Lab works, from a paper to your project.

Agents read the research and build the demos; deterministic gates decide what ships; a retrieval layer matches it to your case. Play each story or step through it.

POCs live
53
gates before production
2
documents in the retrieval store
156
bypasses allowed
0

Every morning an author agent reads the research, writes the demo and opens a pull request. Two independent gates decide whether it ships. Nobody, not even us, can skip them.

sourceResearch feedsGmail · newslettersagentAuthor agentClaude · daily 9:00dataLab memorykb.py · BM25 + RRFservicePull requestGitHub · poc/<slug>gateSmoke testCI · 115 handlersgateIndependent reviewerGroq · 12 rulesserviceProtected deployauto-merge · no bypassserviceLive demosHugging Face SpacedataRetrieval storeSupabase · pgvectorservicevectorridgelabs.comVercel · Lab map
01 / 08

The agent reads the research

At 9:00 Lima the author agent reads the day’s AI research newsletters and lists every paper and announcement.

sourceagentdataservicegate

Design rules

  • Agents propose, gates decide

    The author agent can open a pull request but cannot publish. A smoke test and an independent reviewer must both pass.

  • Rules cannot be edited by the work they judge

    The reviewer runs its code and checklist from the protected branch, never from the pull request.

  • Ground the AI in what exists

    The planner only sees the techniques retrieved for your case, so it cannot recommend what the Lab has not built.

  • Open by default

    Demos are public and the code is MIT licensed, so you can verify before you integrate.