Run a multi-step agent pipeline on GPUs you control
Each step gets the card it needs rather than the one the whole pipeline was sized for. The shape of the work and the cost of the work become the same picture.
Everything below describes what this is being built to do. It needs Studio and Compute, which you can look around today but cannot yet use for the whole of this job. The problems are real now; the answers are the part still being built.
A pipeline sized for its heaviest step pays for that step the whole way through.
Why it is hard
One machine has to fit the largest stage
An embedding pass and a 70B generation do not need the same card, but a single box has to be bought for the bigger one and then runs the smaller one on it.
Idle stages still hold the hardware
While a step waits on an API, a human or a queue, the GPU underneath it is doing nothing and costing exactly the same.
The diagram and the runtime drift
The picture of the pipeline lives in one tool and the thing that runs lives in another, so within a month the diagram is a description of what it used to do.
What answers it
A step holds its own card
Hardware is attached per step rather than per pipeline, so each stage takes what it needs. The canvas for this exists today and you can draw one now.
A step that is waiting holds nothing
The intent is that a card is held for the seconds a step runs, on the same per-second meter as any rental. This is the part that needs execution, which is not built.
The diagram is what runs
One artefact rather than two, which is the only arrangement in which they cannot disagree.
How billing works
per second, to the microeuro · no minimum · no egress fee · idle costs nothing
Metering starts when the container does and stops when it stops. The rate is frozen the moment you match, so a provider re-pricing cannot change a running rental, and an aborted run costs what it used.
What you use from Kracht
Related jobs
- StudioplannedBring a workflow you already wroteLangGraph, LangChain and CrewAI graphs are import sources rather than competitors. The intent is that an existing graph arrives as nodes you can see, not as a rewrite.Read the job
- StudioplannedPut a person in the middle of a pipelineThe run stops at the approval and waits. That is the difference between something you can put in front of a customer and a script somebody watches.Read the job
- StudioplannedSee what an agent run actually did, span by spanEvery tool call, prompt and token recorded against the step that made it, so a bad answer traces back to the exact call responsible.Read the job