Agents that ship.

The Go/Echo engine executes a checkpoint graph. Each node is one of four kinds — prompt (an LLM call), generator (a media model call), distributor (fans a run out into N child pipelines), connector (waits for the children and merges). A pipeline is that graph plus per-node model config, saved as a template and run from one line of input. The engine wires up the state store, calls providers, tracks costs against the model registry, and streams progress events over a WebSocket so the studio can show what is happening while it happens. That last part matters more than it sounds: it is the difference between a system I trust and one I babysit.

Every checkpoint picks its model from the FastAPI model registry. The registry is the platform's one source of truth about which models exist, what request schema they take, and how many credits a call costs. A checkpoint says "give me an image generator with this schema" and the registry answers with a concrete model — gpt-image@2 on Runware for a stylised still, an SDXL LoRA served warm from the Modal GPU bridge when the character has to be consistent, Gemini Flash for planning because it is cheap and structured. Swapping the model for a checkpoint is one config change; the pipeline shape does not move.

Three things. First, structured JSON contracts between checkpoints — an LLM that returns freeform prose is a bug, so every prompt checkpoint declares the shape it expects and the next node parses against it. Second, validator checkpoints — a second, cheaper LLM re-reads the payload and repairs the JSON before the run continues; the carousel pipeline in demo 04 is the smallest version of this pattern, and I use the same shape everywhere. Third, per-checkpoint cost metering — every call is billed against the registry's pricing and drawn from the user's credit balance, so a runaway pipeline hits a ceiling instead of my card.

One piece I am particularly fond of: the character lab. It is a FastAPI service that talks to a Modal GPU bridge. You feed it a folder of reference images; it curates a training set, trains an SDXL or Z-Image LoRA on A100s, and registers the result back into the model registry. From then on, any generator checkpoint in any pipeline can name that LoRA and the character rides consistently into every scene. The reel pipeline in demo 05 is exactly this trick — the seed portrait pins the host down, and twelve independent scene generations still look like the same person.

The platform started as a single-user tool for me and grew into a small multi-user system: logins, per-account credits metered against real model pricing, node-graph generator profiles that let a user configure their own default pipelines. The reason the transition was survivable is that the engine only ever knew about pipelines and credit ledgers — user identity got attached at the studio edge and rode along in the run context. Everything else stayed the same.