Platform / Definition

Agent IO

Specialists woven together on the nQube backbone.

Product screens recreated with synthetic data.

app.nventr.io/orgs/acme/agents
replayLaunch planning swarmExample: Product launch — swarm of swarms
running1/3 roundscap 6 roles0 tok200ms

ObjectiveLaunch the Team plan self-serve in six weeks: pricing page and plan matrix, in-app upgrade with billing, two CRM integrations, a launch announcement, and support readiness.

JudgeCoordinatorround 1 · planning
R2R3
Shared notes
Round 1
Scripted replay of a coordinator run. Names are the Agent IO library; numbers are illustrative. The coordinator names roles from the library, runs them in parallel then in order, takes their suggestions under the cap, and judges each round. It cannot commit — the gate and the write live in the nQube pipeline.
support-intakeenabledRotate token
System prompt
Answer from projects and pipelines. You may propose a pipeline. You may not create a workload. You do not post to external systems.
Can read
projectpipelineworkload
Can propose creating
pipelineprojectworkload
Look
accentrightChat
Install
script · embed.jscopy
iframe · /agent/framecopy
API · sessioncopy
Specialists woven on the nQube backbone
nQube backbone
Inquiry
Product
Build
Explain
control.coordinateCoordinator

Weaves roles on one objective

blocksnQube compute

Shared, dedicated, GPU

control.review / filterThe gate

No model reaches a side effect

declared capabilitySide effect

Webhook, dataset, S3 — named

The library lives here. One role embeds as nventr Agent. Several roles run together on an nQube coordinator. That block cannot commit.

What it is

Agent IO is the catalog of agents for your org. Each card is four steps. Voice is the name, the greeting, and the system prompt. Access is two lists: which records it may read, and which records it may propose creating. Reads you turn on are reads. Creates are proposals. Nothing on that list is created until a person confirms.

Look is the accent, which side the launcher sits on, and the short label on the button. Install, once the agent is enabled, is three copy blocks on the same token: a script for the floating launcher, an iframe for a fixed panel, and the session API. That install is nventr Agent — the thing a visitor actually talks to. Enable, disable, and rotate stay on the card. Rotating the token stops every copy.

The permissions are platform records: workspace, project, pipeline, workload. This is not a connector to your ERP. If the job is “post a vendor invoice into a system you already run,” that job is an nQube workflow on agentOS, not a toggle on this screen.

The library is a set of specialist roles — research, spec, architecture, engineering, QA, security, and the rest of a small company sequence. Adding one creates an agent with that prompt. A coordinator block in nQube is what runs several of them on one objective. The embed of any one of them is still a single nventr Agent conversation.

What it is not

Agent IO is not agentOS. It does not commit workflow writes, and it does not hold your ERP secret. The library defines specialists. Running them together is an nQube coordinator block, and that block still cannot commit. Turning on “can propose creating a pipeline” does not give the agent a finance system. It lets the agent ask a person to create a pipeline record.

Who it is for

The operator who wants a named assistant on an intranet or a product surface, with a boundary written down. Also the person who will be asked which agents exist and what each one can touch.

Where you get it

Platform. Defining agents and embedding them is part of Platform. Production credit for conversations starts on Team.

Pricing →

A support agent with a short leash

Synthetic example. Not a customer.

You name it support-intake, write a prompt that answers from your projects and pipelines, and turn on read for those two. You allow it to propose a new pipeline and nothing else. You set the accent and leave the launcher on the right. You copy the script onto the help page. A teammate asks it to “just create a workload.” That toggle is off, so the agent cannot. A request to draft a pipeline comes back as a proposal with a confirm button. Until someone confirms, no pipeline exists.

What changes when you have it

A written boundary

Read and propose are toggles, not a paragraph in the prompt.

A token you can kill

Disable or rotate without shipping a new page. Every install uses that token.

Three ways onto a page

Script, iframe, or API. All three are nventr Agent, running as this definition.

Inside Agent IO

  • Voice: name, greeting, and system prompt
  • Access: read and propose toggles for workspace, project, pipeline, workload
  • Look: accent, launcher side, and a short launcher label
  • Specialist library for the nQube coordinator
  • Live swarm view: rounds, roles, shared notes, suggestions and verdicts per coordinator run
  • Install: script, iframe, and API from one token

Several roles, one objective

1

The library

Research, spec, architecture, engineering, QA, security, and the rest. Adding one creates an agent with that prompt.

2

The coordinator

An nQube block reads the objective and names roles from this library. It does not invent a role.

3

Together, then in order

Parallel roles run at the same time. A sync role runs after, so it can read what they wrote. A specialist may ask for another library role; it joins while the cap allows.

4

The loop

A judge closes each round. Another round runs until the objective is met, the round cap hits, the token budget hits, or the clock runs out. A failed role is retried once.

5

Watch it live

Every coordinator streams its rounds, roles, notes and verdicts to Agent IO as it runs — including the swarms a sub-pipeline starts. When nothing is running, the page replays a scripted example.

6

Still one conversation

Any single role can also be nventr Agent on a page. That embed does not split itself into specialists.

7

The gate

The coordinator cannot commit. agentOS writes only after a person or a keyword filter, and only to a capability on the pipeline.

How it works

1

Voice

Name, greeting, and system prompt. This is how it speaks. It is not the permission.

2

Access

Choose which platform records it may read, and which creates it may propose. Off means it cannot.

3

Look

Accent, launcher side, and a short label. The widget picks these up from the definition.

4

Install

Enable the agent. Copy the script, the iframe, or the API. Rotate the token and all three stop.

Agent IO. Agent IO decides what may be asked. One agent asks it on a page. Several roles run together on nQube. agentOS is the door that commits a workflow.

The names next to it

See it on your workflow

Platform Start is the self-serve canvas. A first pipeline we stand up with you is delivery, priced apart from the subscription.