This page is about the AI agents that work inside Syllabium, not a course that teaches agentic AI. Syllabium is the learning management system (LMS) your training runs on, and its AI agents work in it beside your people, under the same checks.
Syllabium is an enterprise learning management system (LMS) for regulated industries, made by Athantic AI LLC.
Each agent has authority granted capability by capability, its safety-critical work waits for a person, and every attempt it makes is recorded with the check that decided it.
An AI agent in Syllabium is a named actor with its own account and role, which does defined work in the platform under the authority an administrator grants.
AI agents in Syllabium help author content, watch engagement, check assessment questions, draft notes on reflections and run the platform’s quality checks.
| The agent’s work | What it does | Where a person decides |
|---|---|---|
| Authoring | It works beside authors in the studio, on a block’s text, in Off, Ask, Suggest or Act mode. | In Suggest, the author applies each change; in Act, it writes only the author’s own draft, with undo. |
| Engagement | Each day it reads three counts per learner (overdue lessons, missed reflections and days unseen) and raises a suggested intervention in fixed words. It calls no model and reads no learner’s writing. | Every raise waits for a staff member, who approves, rejects or resolves it. |
| Question checking | A verifier, then a separate critic, checks each assessment question against the slides it cites. | A rejected safety-critical question goes to a reviewer. |
| Reflection notes | It drafts a note for a supervisor on a learner’s reflection, and the learner never sees it. | It runs only when the enrollment consent names it and a person turns it on. |
| Quality assurance | It runs the platform’s built-in test suite and writes only a short summary of each run. | Rules decide every verdict, never a model, and a failed run stays failed, acknowledged by a named person with a reason. |
Further roles, among them a tutor, a grader, a translator and a proctor, are on the roadmap. Block-level LLMs and Agentic RAG describe two kinds of AI work in more detail.
For each AI agent and each capability, an administrator sets one of four outcomes: Allowed, Human review, Admin review or Blocked. Unset means refused.
| Outcome | What happens |
|---|---|
| Allowed | An explicit grant for a named agent and capability. |
| Human review | The action waits for a person. Safety-critical work by an AI agent does too. |
| Admin review | A different administrator must decide. |
| Blocked | The refusal and its check are recorded. |
| Unset | The action is refused. |
Authority also runs in levels, from viewing through drafting and ordinary publishing to safety-critical publishing, deleting and full authority, and above ordinary publishing an agent’s act always waits for a person. A grant can be limited to one library or one block.
AI settings exist at three levels, the platform, each tenant and each organization, and a lower level can only be stricter than the one above it.
The administrators of each tenant and each organization set their own level. A cap across all agents applies at each level, beside each agent’s own limits on actions, tokens and spend. Loosening a safety-critical setting, any change an agent proposes, and a loosened rule each wait for a second, different administrator. Your own models explains how each level’s route chooses the model.
On content flagged safety-critical, an AI agent’s work waits for a person. No setting changes that.
AI review before publish explains the review in full.
Every attempt by an AI agent is recorded with the check that decided it, whether the act was allowed, refused or failed.
Before any agent act, twelve checks run in this order.
| Check | What it settles |
|---|---|
| Emergency stop | Whether a stop is in force for this work. |
| Switches | Whether this kind of AI work is switched on. |
| Tenancy | Whether the act stays inside the right tenant and organization. |
| The role’s state | Whether the agent’s role is active. |
| Permission | Whether the agent holds this capability. |
| Safety-critical rule | Whether the content is safety-critical, in which case the work waits for a person. |
| Override | Whether an override applies. |
| Level asked | Whether the grant reaches the level of authority the act needs. |
| Library scope | Whether the grant covers this library. |
| Limits | Whether the act is within its rate and spend limits. |
| Data class | Whether this kind of data may be sent to this model’s provider. |
| The decision | Allowed, Human review, Admin review or Blocked. |
A refusal names its check, is recorded, and never reaches a model. Every model call is logged with its model, tokens and cost, in tamper-evident records kept seven years. The platform’s quality suite tests that each of the twelve checks can refuse. Evidence explains how the records are chained and checked.
People and AI agents act through one route: the same authority check, then the act, then a record, written whether the act was allowed, refused or failed.
An agent’s action takes the same path a person’s does, and no service has a back door for AI.
| People | AI agents | |
|---|---|---|
| Identity | Their own sign-in. | Their own account, with no password and no sign-in. |
| Authority | Roles and capability grants, each with a reason. | Capability grants per agent, each with a reason and an outcome. |
| Safety-critical content | A second person approves each change, and an administrator can switch that off for people’s own acts, with a second administrator. | The work always waits for a person, and no setting changes that. |
| The record | Every act is recorded. | Every attempt is recorded with the check that decided it, and every model call with its model, tokens and cost. |
An emergency stop, read before every agent act, can halt all AI, one agent, one kind of work or one type of content.
It is set in two steps, with STOP typed again to confirm, and a red banner shows on every AI page while it holds. Each agent also stops at its limits on actions and tokens an hour, and on spend a day and a month. Governed AI shows the one route every AI action passes, and Platform the rest of the system.
Syllabium keeps a record that reviewers can read: who each agent is, what it may do, what waited for a person, and which check decided each attempt.
An administrator exports a period’s records as one archive, with a verification report and a list of file fingerprints, and every export is logged. Evidence explains the export for an auditor.
Enroll your organization online, see the pricing, or talk to us. Syllabium is made and operated by Athantic AI LLC.
Contact us