Under the hood

The engine, and the apps that snap into it

There is exactly one engine. It is domain-agnostic: a step scheduler with retries and checkpoints, a fleet of scrapers and fetchers, a knowledge index, and a cost meter. It knows nothing about hotels or helmets. Everything domain-specific — a channel manager, a research pipeline, a CAD project — is an app that snaps into the engine through standard sockets.

This is why a new project in a new domain isn't a rewrite. The engine is already battle-tested; the app is a thin layer of domain knowledge plus the parts it pulls from the catalogue.

How the router classifies what you drop

The chat box is a router, not a chatbot. Anything you type or paste — a sentence, a URL, a repo, a PDF — is classified along two axes: intent (build / enrich / research / chat) and kind (component / skill / practice / idea / question). A pasted skill link resolves to its SKILL.md, gets a maturity guess, and is proposed as a diff into the staging inbox. The router never commits to the brain on its own — it stages and proposes, you approve.

How learning loops write back

Every run is instrumented. When a part survives a project without new bugs, the close-out ritual proposes a maturity bump. When the agent solves something genuinely new, it proposes extracting a skill. When something breaks and gets fixed, it proposes documenting the failure mode. These are proposals — auto-generated, human-approved, written as brain diffs. The brain is versioned; a bad write is one revert away.

Auto-propose, never auto-commit. The moat is a vetted library, and vetting is the human's job.

Scheduled research sweeps

Parked ideas in the garden each carry a cadence. On schedule, a sweep runs targeted research against the open question that's blocking the idea — a materials constraint, a regulatory threshold, a competitor's launch. Most sweeps find nothing and quietly reschedule. When one finds something that changes the calculus, the idea card wakes up with a summary and a one-click “spin into project.”

The cost model

Cost is shown before you commit, never revealed after. The estimate is a function of plan length, autonomy level, and which steps invoke models versus reused parts. Reused battle-tested parts are nearly free — they don't re-run inference. The autonomy dial is the honest knob: Guided keeps you in the loop and cheap; Max removes the brakes and shows you the bill it implies before you pull it.

Domain-pluggable verification

“Done” always means demonstrated. The verification bus accepts pluggable backends per domain: for software it drives a headless browser and narrates the clicks; for an engineering project it runs a simulation pass and looks up the governing standard (e.g. EN1078 impact requirements) and extracts the binding constraints. The studio never says “fixed ✓” without showing the evidence that earned the checkmark.