Can our own developers build here, or is it low-code only?

Both. Business teams build on the visual canvas; engineers import their own services from an OpenAPI document as steps, and call published agents through scoped API tokens. Every agent is governed the same way, whoever built it.

What happens when a step fails?

Every run is recorded step by step, so you can see which step failed and what it was given. A flow that gets worse is caught by its evaluation set before a new release can go live.

Do we need the whole platform to use AI Factory?

No. Most teams start with one template on one knowledge source and grow from there. The Factory runs on the Brain, so the same permissions and the same audit apply as you add more.

SphereIQ AI Factory — Agents Built as Flows, Tested Safely, Released With a Review

What is SphereIQ AI Factory?

AI Factory is where agents are built as flows you can read: steps that retrieve, decide, act and hand a decision to a person — tested with every change simulated, graded by evaluations that gate a release, and released through a review that leaves a signed receipt.

A factory has stations. Here are the six.

Design
Flows on a canvas
An agent is a flow you can read: steps that ask, retrieve, decide, act, loop and hand a decision to a person, each configured on its card. Engineers can import any service from its OpenAPI document as steps.
Start
Four ways to begin
A person asks, a schedule fires, another system sends an event, or a file arrives — so a digest, an intake queue and a document review each start the way they really do.
Test
Changes simulated
Press Test and the flow runs: reads happen for real, and anything that would change a record or send a message is simulated and shown, step by step.
Evaluate
Evaluations that gate
Question sets graded against what the flow produced, with pass rate, latency and cost per run. A gating set decides whether a release can go live; a question that could not run leaves it unproven, not passed.
Release
A reviewed release
Publishing shows what changed against the live version, re-runs the test questions at request and again at decision, and on a production deployment needs someone other than the requester to approve. Each release gets a signed receipt.
Operate
Output and run history
Scheduled and event runs land in an Output queue with assignment, review state and feedback. Every run is recorded step by step.

The Output queue in the product: what each agent produced, the agent, who it is assigned to, notes, the cost, the outcome and whether it has been reviewed. One run is waiting on a person; two are waiting for approval in all, and all 22 are not yet reviewed. — Output in the product: what each agent produced, who it is assigned to, what it cost, how it finished and whether a person has reviewed it — one run is waiting on a person. A demo workspace.

Anatomy of an agent

A draft stays a draft until a release is approved.

Draft agent

Each model step carries its own instructions and its own model.

Retrieval applies each document’s permissions before anything reaches the model.

OpenAI or Anthropic, chosen step by step.

The questions it is checked against before a release goes live.

The Skills list in the product: shared instructions such as a house report format, an incident report format, numbers exactly as written and saying plainly when it does not know, each with its category, version and whether it is built in or the workspace's own. — Skills in the product: shared, versioned instructions any step can use — a house report format, a rule about numbers, when to say it does not know. Changing one bumps its version. A demo workspace.

The release gate

A release cannot go live until its gating test set passes — and on a production deployment, the person who asked for it cannot be the one who approves it.

Drafttest set · pass rate ≥ thresholdLiveA refused release stays in the release history, with its reason.

Drafttest set · pass rate ≥ thresholdLive

A refused release stays in the release history, with its reason.

Flow evaluations report pass rate, p50 and p95 latency, and cost for every run. Retrieval is scored separately, by citation hit-rate against ground-truth answers.

Signed

Every release carries an Ed25519-signed receipt that can be verified without us.

Two people

On a production deployment, the person who requests a release cannot approve it.

Unproven

A test question that could not run makes the gate unproven — never a pass.

An evaluation set in the product, named Plan documents: answers with sources. It passes at 80 percent, gates releases and last scored 100 percent. Two checks: the answer cites at least one source, and it never contains I don't know. Below it, the questions it asks, such as how the vesting schedule works and how a member applies for a hardship withdrawal. — An evaluation set in the product: the questions an agent is checked against, and the checks on every answer — cite at least one source, never say “I don’t know”. It gates releases, so a new version cannot go live until it passes. A demo workspace.

Start from one of 60 templates.

Each one is a working flow — steps, sources, checks and approvals already wired — including templates for pension and retirement funds.

Every template opens in the builder as the flow drawn on its page.

Browse all 60 templates

Starting a new agent in the product: a dialog titled What are you building?, with kinds of agent grouped under Knowledge and Documents — answers questions, researches a question (marked partly built), pulls a record together, watches a web page, writes a document, fills in a questionnaire, pulls data out of documents, compares two documents, reviews against a standard, translates — each with how many steps it starts from. — Starting an agent in the product: pick the kind of work first, and it opens as a flow with its steps laid out. One kind is marked partly built, because it is. A demo workspace.

Governed actions

When an agent's output triggers an action — an email, a Jira ticket, a Slack message, a row written back to a system of record — it passes five checks before anything leaves this deployment.

01
Validation
The arguments are checked against the action's own schema before anything is called.
02
Firewall
The payload is scanned for injected instructions, keys and personal data.
03
Policy
Per action and per environment: run it, hold it for a person, or refuse it.
04
Approvals queue
A held write waits with what approving will do, and who it is waiting for.
05
Execution
It runs once. A write is keyed by run, step and call, so a resumed run never repeats it.

Every step audited.

Asked before every Factory rollout.

Can our own developers build here, or is it low-code only?: Both. Business teams build on the visual canvas; engineers import their own services from an OpenAPI document as steps, and call published agents through scoped API tokens. Every agent is governed the same way, whoever built it.
What happens when a step fails?: Every run is recorded step by step, so you can see which step failed and what it was given. A flow that gets worse is caught by its evaluation set before a new release can go live.
Do we need the whole platform to use AI Factory?: No. Most teams start with one template on one knowledge source and grow from there. The Factory runs on the Brain, so the same permissions and the same audit apply as you add more.

A workforce this capable needs exactly one thing above it.

Chapter 05: the reason your auditors, CISO, and CFO will say yes.

Book a meeting Final chapter: Governance

Contact us

From one agent to an AI workforce.

What is SphereIQ AI Factory?

A factory has stations. Here are the six.

Flows on a canvas

Four ways to begin

Changes simulated

Evaluations that gate

A reviewed release

Output and run history