Capture · Replay · Distill · Evaluate — the layer that plugs into any harness you already use and makes it get better over time. The durable value is the loop, not the orchestration.
Developers → a provenance-first registry · Teams → a managed loop
Frontier-only tooling prices out individual developers — and for many teams, the code simply can’t leave its own environment. The improvement should run against the tools you already have, with the option to keep everything local.
Every successful coding episode becomes fuel for the next one. Capture what happened, replay what worked, distill it into reusable skills and small models, then evaluate the result — and feed it back in.
real coding episodes, as they happen
the trajectories that actually worked
reusable skills + small models
under a controlled protocol
↺ the output of evaluate feeds the next capture — it is a loop, not a pipeline
Plugs into the harnesses you already use
From the harnesses you run, up through the loop, evaluation and provenance, the registry, and — specified but not yet built — a managed cloud control plane.
hosted memory · execution · metering · gateway
spec'd, not builtpublish · verify · rank — only verifiable results listed
provenance-firstmulti-mode runner · CL metrics · promotion gate · manifests
capture → replay → distill
thin connectors into the coding harnesses you already run
a conceptual map — each layer is a responsibility, not a repository
Publish small fine-tuned coding agents and reusable skills. Rankings are provenance-first: a result is listed only if its manifest and digests verify. No verifiable manifest, no entry.
The loop, hosted: managed memory, execution, metering and a gateway — with code-can’t-leave treated as a hard constraint, not a tier. Specified in detail; built once demand is proven.
We treat continual learning as a claim to be proven under a controlled-exposure protocol — not a result we are announcing. The metric definitions live in code as the source of truth; what follows are the methodology and the success thresholds, not numbers we have already hit.
Does learning earlier tasks help on later, unseen ones?
Success threshold: FWT > 0
Does new learning erode what the agent already knew?
Success threshold: BWT ≥ −ε (bounded forgetting)
Can the agent still absorb genuinely new skills over time?
Success threshold: Plasticity ≥ 0.5
The protocol, in brief
Same student model and command budget across arms; a pinned benchmark and dataset hash; a frozen zero-shot baseline that runs ephemeral with no memory leakage; fixed or repeated-randomized task order reported with variance. Numbers reach this page only once a manifest-backed entry exists — until then, thresholds and method.
Early and bootstrapped. Tell us whether you’re a developer or a team and we’ll reach out as the loop opens up.