LoopClimber creates candidate workflows, runs deterministic tests with quantified tolerances, prunes slower or failing graphs, and promotes the fastest passing workflow through an outer meta-learning loop.
The AI workflow node becomes a container. A large frontier model proposes prompts, retrieval, tools, handoffs, and smaller execution models; every proposal is saved as a candidate sub-DAG.
The container unfolds into inputs → large-model proposal → context filter → candidate graph → human checkpoint → distilled small model. Every inner-loop node stays explicit and replaceable.
The execution node fans into code that checks quality ≥95%, calibration error ≤0.5%, tool errors <0.5%, p95 wall clock <4s, cost <$0.35, and unsafe actions =0. Every branch must pass its tolerance.
The outer loop sends matched cases through candidate sub-DAGs. Edge rates show throughput; the wall-clock edge slows, blocks red, queues work, clears, and returns to blue.
The outer loop updates metric tags from its meta-learning record, nests new recovery evals under promising nodes, deletes failed branches, and distills successful behavior from large models into smaller, faster models.
Start an expiring practice room, operate the real Durable Object API, and share its link. Conflicting edits rebase once, so collaborators converge on one revision.
The surviving inner loop becomes the next baseline. Its traces and failures expand the regression bank, so the next outer-loop generation searches a smaller graph with stricter tests and lower wall-clock time. Read the research outline.