Skip to main content

​
Purpose

Gradientflo is the second application on Smaya, and it runs a workforce of 10 specialized agents. Each is accountable for one job. You combine them into AI Employees — named roles with a mandate, bounded authority and named human owners. Most of them teach. Onboarding and project training are delivered by AI tutors — conversational agents that carry your organization’s project, team and company knowledge to a new hire and answer, explain and coach through the first ninety days. They run over the same voice and avatar interfaces as Exterview’s interview agents, so a new hire talks to them rather than reading a portal.
Availability. Available means you can use it today. Upcoming means it is on the way and not yet usable. Gradientflo is a separate, Upcoming application, so every Gradientflo agent is Upcoming. marks the seven that seed first.

​
Onboarding — 4 agents

The first ninety days, where most of the value of a good hire is won or lost.

​
Development — 3 agents

Closing the gap between the person hired and the person the role needs.

​
Performance — 2 agents

Whether the hire is working, and what to do when it is not.

​
Operations — 1 agent

The module’s first write-capable agent. It is the row most likely to face a security review rather than pass one, so it ships behind permission scoping and human confirmation on every write — not ahead of them. Why these seven seed first. They are the closed loop: a hire is answered, checked on, watched for risk, and re-evaluated against the criteria they were hired on. The other three extend that loop outward — into learning content, and into acting on your systems — and both extensions depend on the loop working first.

​
When a tutor does not know

A design decision worth stating plainly, because it is unusual and deliberate. The tutors answer without escalating to a human. A new hire’s whole reason for asking an AI tutor is that they do not want to interrupt a colleague — or do not want to look like they should already know. An agent that answers “ask your manager” hands the problem straight back and defeats its own purpose. So instead of escalating, a tutor that is not confident enough to answer defers and logs: it says it does not have a reliable answer yet, records the gap, and the gap becomes a signal that your onboarding material is incomplete.

Never a confident guess

Answers come from your own curated material with citations. Below the confidence threshold, it declines rather than improvising.

Never 'ask your manager'

Deferring back to a human is the failure mode this agent exists to remove, not its fallback.

The gap becomes the signal

Every deferral is logged. A cluster of them tells you exactly which part of your onboarding material is missing.

Writes still need a person

This applies to answering only. Anything that changes a record or contacts someone still waits for a named human.

​
What no agent may ever do

Judge performance

Role fit and development output is evidence prepared for a manager. A judgment about a person’s performance is made by a named human.

Assess without notice

No agent runs a background assessment of an employee before that employee has been told it happens, and the notice is recorded as its own act.

Exceed the person using it

Effective authority is the intersection of the agent’s scope and the user’s permissions.

Act without a record

Every action, proposal and approval is captured and attributed.

​
Why this proves the platform

Three of the ten run on the same underlying expertise as a recruiting agent — the Knowledge Agent, the Journey Nudge Agent and the Resume Screening Agent each now serve two business functions. That reuse is the entire argument for an Agent OS. A second function does not mean a second product: it means new roles composed largely from capability that already exists, already governed, already audited.

The subject changes

Recruiting agents act on a candidate. HR agents act on an employee. The platform treats who is being assessed as a declared property, not an assumption baked into the code.

The scope changes

Recruiting work is scoped to a requisition. Onboarding work is scoped to an onboarding journey. The rule that criteria never change mid-stream applies identically to both.