
Every agent picks its own model
An AI agent calls whatever model its developer typed into the code. Nobody at the company decided that, and changing it means editing the agent.
- Unreviewed models read your data. A premium model nobody approved can receive customer records at premium prices.
- Models retire on a schedule. Anthropic retires older models and says requests to a retired model fail.¹ OpenAI says a shut-down model no longer accepts requests through its standard API.² An agent that hard-coded the name stops working.
- One provider is a single point of failure. When it has an outage, every agent that depends on it stops too.
- Access is the gap. IBM found 92% of organizations with an AI-related breach lacked proper AI access controls, and incidents involving unapproved ("shadow") AI reached 43% of the breached organizations IBM studied, up from 20%.³
The question is not which model is best. It is who decides, where, and how you prove it.
One approved list, enforced at the gateway
Argent Strategic Intelligence sits between your agents and the models they call. No model runs until an admin approves it.
- Default deny. An empty list allows nothing. A model that is not on the list is refused at the gateway, so the request never reaches it.
- Three layers, each narrower. The list is platform AND tenant AND agent. Each layer can only narrow the one above, so delegation never widens access, and a refusal names the layer that said no.
- Approve by provider and name, with wildcards, for example every model in one approved family.
- Decided on the real request. The gateway reads the model from the request itself; a request whose model cannot be read is refused, not let through.
- Finer rules when you need them. Model and endpoint access rules allow or deny by provider, route type, model, hosted tools, function tools and output type, for the whole tenant, a department, a person or one agent. The narrowest scope wins and a locked deny always wins.
This covers every agent that works through Argent: coding agents, support assistants and copilots alike. See securing AI coding agents, governing customer-facing AI and rolling out copilots safely.
Model Swap: run the approved model instead
Refusing work is not the goal. A Model Swap rule turns a request for one model into a request for an approved one, without touching the agent.
- The agent asks for model X. Often a premium model, exactly as its code says.
- A platform rule matches. "When an agent asks for X, run Y." Rules are an ordered list; the first one that matches wins. Only a platform administrator writes them.
- Argent checks the target. Y must be approved, priced and reachable with a route and credential. The rule fires Always, or only when a budget reaches a set level (80% by default).
- The request runs as Y. At the same provider only the model name changes. At another provider Argent translates the request there and the answer back, so the agent keeps its own format, streams and tool calls included. Proven live with Claude Code and Codex CLI across OpenAI, Anthropic, Google Gemini, AWS Bedrock and Azure OpenAI.
- The answer is honest. The reply names the model that really ran, and the signed record shows the model asked, the model that ran, the rule and an estimated saving for that call.
Owners choose what happens when a swap cannot run: Fall back (run the asked model only if it is approved itself), Always deny, or Open (leave the choice to the level below). A higher level, such as the tenant or a department, can lock that choice for everyone below it.
Model classes: agents ask by role, not by name
A model class is a name an admin picks, such as `fast`, for an ordered list of approved models. Agents put the class name in their model field, and Argent picks the model.
- Change once, every agent follows. Point
fastat a new approved model in one place, and every agent that asks forfastmoves with it. No agent changes a line of code. - Failover in order, by default. If a model fails before it answers, the request moves to the next approved model: another provider, or another endpoint for the same model. Every check runs again on each hop: policy, credential, budget, data protection and the audit record.
- You define failure. Connection errors, server errors, rate limits, timeouts, a slow first token, model not found, or an overloaded provider.
- Or spread agents by weight. Sticky mode keeps each agent on one member, split by the weights you set.
- Classes only hold approved models. A member that stops being approved, priced or healthy is skipped, and the record says why.
- Budget-aware. A swap rule can point at a class: at 80% of budget, move the agent to
fast. At 100%, the budget still stops it. Spend is metered at the price of the model that ran. Pairs with AI spend caps.
Classes are granted to the tenant, a team, a person or one agent. Creating or deleting a class, or changing the names agents use to reach it, takes a passkey step-up, and each agent's Connect tab lists the class names it may use.
What the operator sees
Every model decision lands on a signed record, and the console shows it in plain words, with the reason code beside the English.
| Where | What it shows |
|---|---|
| Allowed models (on the PDP config page, the platform policy settings) | The platform approved list. An empty list warns that it blocks every model call. |
| Model and endpoint access | The tenant list, access rules, Off / Monitor / Enforce, and the impact of a draft rule on the last 7 days of traffic. Each agent's own list sits on its Access tab. |
| Model Swap rules | The ordered rules, each "asked model -> approved model", Always or Budget at a set percent. A target that is not approved cannot be saved. |
| Model classes | Each class, its members in order, failover or sticky, and how many members are usable right now. |
| Request inspector | "Ran <model> instead of <model> (rule 1), est. saved $x" or "Routed through class fast, ran <model>, after 1 failed hop". |
| Signed record | Agent, model asked, model that ran, rule or class, hops, and the reason code, such as llm.deny.model or llm.allow.model_swap. |
When an examiner asks which model read a customer's complaint, the record answers. Go deeper with session replay and one-click audit evidence.
Where it fits, and how you start
Model Control serves the goals that come up first when AI reaches production.
- Goals: it is the core of model control and routing, and it feeds AI budget and token spend, audit and accountability and compliance evidence.
- Industries: in financial services, insurance, healthcare, the public sector and retail, it keeps customer data on models you reviewed. All industries.
- Regulated and sovereign AI: approved models can include ones you host yourself, inside your own network. See AI in regulated and air-gapped environments, sovereign AI and deployment options.
- Find what is in use first with shadow AI discovery, then approve it.
- Watch. Argent records what it would refuse or swap, and refuses or swaps nothing until you switch on Enforce. See watch before you enforce.
- Approve. Build the approved list, the first swap rules and a class or two, then switch on Enforce.
- Prove. Every request's record shows the model asked, the model that ran, and why.