Model Lifecycle Management
ANTS Platform helps you govern which third party LLMs and agent frameworks your agents use, and see when a model is reaching end of life. It covers four things:
- A catalog of LLMs and agent frameworks with lifecycle status and dates.
- Per-organization approval of models and frameworks.
- Detection of unapproved models in your traces.
- A governance profile for each agent, with a deployment stage and a review workflow.
The AI Inventory and agent governance profiles are part of the governance features available on the Enterprise plan.
Model and framework catalog
Open AI Inventory in the app to browse the catalog. It has two tabs: LLM models and agent frameworks. ANTS maintains the catalog entries.
Each LLM model entry can include:
| Field | Example |
|---|---|
| Identity | Name, manufacturer, family, version, parameter count, release date |
| Lifecycle status | current, deprecated or archived |
| Lifecycle dates | End-of-life (EOL) date and end-of-support (EOS) date |
| Lifecycle policy | Link to the provider's lifecycle or deprecation page |
| Licensing | License and whether weights are open |
| Technical | Modalities, context window, hosting options, capabilities, training cutoff |
Status meanings:
| Status | Meaning |
|---|---|
current | Available and recommended |
deprecated | Still available but not recommended |
archived | No longer available |
Each framework entry has a type (orchestration, agent or evaluation), supported languages, license, links to docs and source, and the same current, deprecated or archived status with release, EOL and EOS dates.
You can search and filter the catalog by status, approval state and other attributes, and compare models side by side.
Per-organization approvals
Approvals are set per organization. Each model and each framework is either approved or not approved for your organization. You can approve one item at a time or approve and unapprove several items in bulk. ANTS records who approved an item and when. One organization's approvals do not affect another's.
Detecting unapproved models
When an agent's traces contain an LLM generation, ANTS checks the model name against your organization's approvals. Any model that is not approved for your organization is flagged, including models that are not in the catalog.
Where it shows up:
- AI Agent Center. Agents with violations show the time of their last violation, and you can filter the agent list to agents with or without violations.
- Agent Risk tab. The Unapproved Models section lists the unapproved models the agent used and the number of traces, for a time range of up to 7 days. View Traces opens the matching traces.
- Notifications. From the same section you can send a notification to a connected Slack channel, Jira project or Linear team.
Approval changes are picked up within a few minutes. Unapproved model usage also feeds the agent's composite risk score.
Detection flags usage after the fact. Approvals do not block calls to unapproved models.
Agent governance profiles
Each agent can have a governance profile, built on the Evidence Collection page. A profile records:
- Use case and scope, including the agent's deployment stage: Development, Staging, Pilot / Beta, Production or Decommissioned.
- Risk level: Unacceptable, High, Limited, Minimal or Not classified.
- Frameworks the agent is assessed against: EU AI Act, ISO/IEC 42001 and NIST AI RMF.
- Evidence in eight categories: use case and scope, risk classification, data governance, human oversight, security and abuse prevention, bias and fairness, transparency and user rights, and legal and accountability, with attestations.
A profile has one of these statuses:
| Status | Meaning |
|---|---|
| Draft | Evidence is being collected |
| Pending review | Evidence was submitted for review |
| Changes requested | The reviewer sent the profile back with feedback |
| Approved | The reviewer approved the evidence |
| Archived | The profile is no longer active |
A reviewer can send an evidence request for an agent to a teammate, optionally through Jira, Slack or Linear. The teammate accepts the request, fills in the evidence and submits it, which moves the profile to Pending review. The reviewer then approves it or requests changes.
The deployment stage is a recorded field. There is one review workflow per profile, not a separate approval for each stage, and ANTS does not deploy models or block deployments based on profile status.
Planned capabilities are listed on the roadmap.
Related
- AI Inventory
- LLM-as-a-Judge: evaluate models and prompts before you switch.
- Experiments: compare models on a dataset.