
ARGENT FORMODEL CONTROL AND ROUTING
Master Control System · by Nyedis
Your goal: every agent on a model you chose
If your goal is to decide which AI models your agents use, start from how it works today: each agent calls whatever model its developer typed into the code. Nobody at the company decided that, and changing it means editing the agent.
- Companies already run many models. In an a16z survey of 100 large-company technology leaders, 81% use three or more model families in testing or production, up from 68% less than a year earlier.¹
- Models retire on a schedule. Anthropic says requests to a retired model fail.² OpenAI says a shut-down model no longer accepts requests through its standard API.³ An agent that hard-coded the name stops working.
- Access is the gap. IBM found 92% of organizations with an AI-related breach lacked proper AI access controls.⁴
- One provider is a single point of failure. When it has an outage, every agent that depends on it stops too.
The goal is not to pick the best model once. It is to decide, in one place, which models may run, where each request goes, and to prove it afterwards.
Five questions model control answers
Model control and routing is a straight answer to five questions, for every agent that works through the Argent gateway.
| Question | How Argent answers it |
|---|---|
| Which models may this agent use? | One approved list, layered platform AND tenant AND agent. Each layer can only narrow the one above, and an empty list allows nothing. |
| What may it do with them? | Model and endpoint access rules allow or deny by provider, route type, model, hosted tools, function tools and output type, for the tenant, a department, a person or one agent. |
| Where does this request go? | A Model Swap rule can run an approved model in place of the one asked for, and a model class picks from an ordered list of approved models. |
| What happens when a model fails or retires? | A model class moves the request to its next approved model, and one change to the class moves every agent that uses it. |
| Which model really ran, and what did it cost? | The reply names the model that ran. The signed record shows the model asked, the model that ran, the rule or class and the reason, and cost is metered at the price of the model that ran. |
The full detail is on the Model Control capability page.
The path: from watching to routing
Each step adds control, and nothing changes for your agents until you decide.
- See which models are in use. Shadow AI Discovery finds the AI tools and agents at work, and each registered agent's Models tab lists every model it calls, with requests, tokens and cost.
- Connect your providers once. One action in the console creates a provider endpoint with its credential. The provider key stays in your vault, and agents that work through Argent carry only their own Argent key. See Keyless Agents.
- Watch. Argent records what it would refuse or swap, and nothing changes on the wire. Access rules show a draft's impact on the last 7 days of your own traffic before you switch them on. See Watch Before You Enforce.
- Approve. Build the approved list, the first Model Swap rules and a class or two, such as
fast. A Model Swap target that is not approved cannot be saved. - Enforce. Switch on Enforce. From then on a model nobody approved is refused at the gateway, before anything leaves, and swaps and classes route each request to an approved model.
How one request is routed, step by step
An engineer's coding agent asks for a premium model to review a pull request. Its developer hard-coded that name.
- The agent sends the request to the Argent gateway, with its own Argent key, exactly as its code says.
- The gateway reads the model from the request itself and checks it before the request leaves.
- The approved list and access rules decide. If the model is not approved and no Model Swap rule covers it, the request is refused, and the refusal names the layer that said no.
- A Model Swap rule matches. The agent asked for a model that is not approved, and the rule sends it to one that is: "When an agent asks for X, run Y." Rules are an ordered list written by a platform administrator; the first match wins. A rule can fire Always, or only when a budget reaches a set level (80% by default).
- Argent runs the approved model. 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. The provider key comes from your vault at this step.
- If the swap points at a model class and that model fails before it answers, the class moves the request to its next approved model, and every check runs again: policy, credential, budget, data protection and the audit record.
- 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 or class, each hop and an estimated saving for that call.
The agent's code never changed. Model Swap is proven live with real Claude Code and Codex CLI sessions, with requests served by OpenAI, Anthropic, Google Gemini, Amazon Bedrock and Azure OpenAI.
Providers, and models in your own network
Approved models can be hosted by a provider, or run by you inside your own network. The same rules apply to both.
| Where the model runs | Examples Argent governs |
|---|---|
| Hosted providers | OpenAI, Anthropic, Google Gemini, Amazon Bedrock, Azure OpenAI and Mistral, each with its credential brokered by Argent, never held by the agent. |
| Your own network | Ollama, vLLM, LM Studio, NVIDIA NIM and Hugging Face TGI: governed and metered like any hosted model. |
| Anything else that speaks OpenAI's format | Any OpenAI-compatible endpoint, governed with a brokered key. |
| An AI gateway you already run | Traffic that already flows through Portkey, Kong AI Gateway or Cloudflare AI Gateway is governed too. |
Argent speaks each provider's own format: OpenAI chat and responses, Anthropic messages, Gemini and Bedrock, streams and tool calls included. To keep data inside your walls, give teams that handle sensitive work an approved list of models you host, and see sovereign AI and AI in regulated and air-gapped environments. Browse our integrations.
Cost and proof, on every request
Choosing the model is also choosing the price. Proving which model ran is what an audit asks for.
- Gartner predicts AI inference costs per agentic workflow will rise more than five times through 2028, as advanced reasoning models typically use five to 30 times more tokens than basic chatbots.⁵
| What | Where you see it |
|---|---|
| Cost by model | The Spend page breaks dollars down by agent, owner, model and provider. Each agent's Models tab shows tokens and cost per model, with findings such as a cheaper model. |
| Budget-aware routing | At 80% of budget a Model Swap rule can move the agent to an approved lower-cost model or class. At 100% the budget still stops it. See AI Spend Caps. |
| The 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. |
| The request itself | Replay the prompt, the response and each tool call. See Session Replay. |
What frameworks already ask for (NIST AI Risk Management Framework):
- GOVERN 6.1: "Policies and procedures are in place that address AI risks associated with third-party entities, including risks of infringement of a third party's intellectual property or other rights."⁶ The approved list and access rules are that policy, enforced at the gateway.
- GOVERN 6.2: "Contingency processes are in place to handle failures or incidents in third-party data or AI systems deemed to be high-risk."⁶ Model classes move a failed request to the next approved model, and the record shows each hop.
Package it for an examiner with One-Click Audit Evidence, or take it to the board with Give the Board What It Wants.
Where to go next
Model control works best beside the goals around it.
- Capabilities on this path: Model Control, AI Spend Caps, Keyless Agents, Watch Before You Enforce, Shadow AI Discovery, Sensitive Data Redaction and Session Replay. All capabilities.
- Related goals: see all AI in your company, budget and token spend, data loss prevention, audit and accountability, agent identity and ownership, threat defense, compliance evidence and sovereign AI. All goals.
- Use cases: Securing AI coding agents, Governing customer-facing AI assistants, Rolling out copilots and chat safely, Controlling agent tool access and AI in regulated and air-gapped environments.
- Industries: Financial Services, Insurance, Healthcare, Public Sector, Retail and all industries.
- Where it runs: inside your own environment: air-gapped on-premises, private cloud or your public cloud VPC. Compare deployment options.
- See it yourself: take the walkthrough or download the data sheets.