Platform / Design surface
Design the workflow once. Static steps and model steps share the same canvas.
Product screens recreated with synthetic data.
nQube is where the work is drawn. You place blocks, wire them, and decide which steps a model may take and which steps only a person or a keyword rule may pass. The same design runs as a batch over a pile of documents or stays up and takes documents as they arrive.
Analysts drag blocks. Developers drop a container or a template into the same graph. There is not a second product for “the technical version.” Credentials are not pasted into a block. Data is a Listen source the block refers to.
A design that never meets agentOS is a flowchart. nQube is the place you declare the boundaries the runtime will refuse to cross.
nQube does not post to your ERP by itself, and it is not the embeddable chat agent. Drawing the graph is the job here. A coordinator block runs Agent IO specialists and still cannot commit. Committing the write is agentOS. Putting a widget on your site is nventr Agent, one conversation, defined in Agent IO.
The person who owns the process — finance ops, support ops, a platform team — and the person who will maintain the graph. You do not need a separate integration team to see the path from document to write.
Platform. The shared canvas is on every Platform tier. Production connectors and workflow runs start on Start.
Pricing →Synthetic example. Not a customer.
You duplicate the “Product launch” example and swap in your brief. A coordinator plans the launch with six library roles over three rounds. “Call pipeline … each” hands every workstream to a callee that runs its own coordinator and a refine loop. A keyword filter catches any brief that graded itself RISK: HIGH and routes it to the launch lead; the rest flow straight through; both converge. One exec sign-off gates the publish and the announcement. Every coordinator in the run is visible live in Agent IO, and the block lights show the recursion as it happens.
Batch and streaming are modes of the same graph, not two projects.
Gates and capabilities sit on the canvas, where an auditor can see them.
Each block shows queued, processing, done, or failed, with the last output beside it. Coordinators stream their swarm to Agent IO.
Every loop and every self-call has a cap on the canvas and a hard stop in the runtime. A model is never the only thing that decides to stop.
control.coordinate reads the objective and names roles from the Agent IO library. It does not invent a role that is not in that library.
Research and security run together. Each one returns text, and may suggest another library role.
Spec runs after they finish, so it can read what they wrote.
A judge checks the objective. Another round runs until it is met, the round cap hits, the token budget hits, or the clock runs out. A failed role is marked failed and retried once.
control.review sits on the path. A keyword filter can sit there instead. A model is not a gate.
agentOS performs the write only after that gate, and only to a capability declared on the pipeline.
Drag inputs, transforms, filters, human review, and outputs onto the canvas. A model step and a template step are both just blocks.
Point a reader at a Listen dataset or connection. The workflow does not carry its own copy of the password.
Each output names a capability you declared: a bucket prefix, a dataset, a knowledge audience, an https endpoint, or a container image prefix. Anything else is not callable.
Batch for a backlog. Streaming when documents keep arriving. The same gates apply either way.
Platform Start is the self-serve canvas. A first pipeline we stand up with you is delivery, priced apart from the subscription.