What Multi-Vendor Routing Actually Means
A custom AI agent does not need one foundation model. It needs the right model for each task it runs.
Classifying an inbound email by type? Cheap, fast, small. A small open-weight model handles it for a fraction of a cent.
Deciding whether to escalate a customer complaint to a human? Higher-stakes reasoning. A frontier model like Claude Opus or GPT-class is worth the cost on this call.
Generating a phone-call voice response? Different infrastructure entirely - a speech model, a low-latency provider.
Summarizing a long support thread? Mid-range, with a large context window.
A serious agent runs hundreds or thousands of these decisions per day, each with a different ideal model. Routing per-task across vendors and tiers is how you get the right answer at the right price, every time. Pinning the whole agent to one vendor's frontier model means you pay frontier-model prices on every task, including the many tasks that do not need frontier intelligence.
This is what multi-vendor routing means in practice. Not a single 'best model' but a portfolio with a router on top.
Who Locks You In, and How
Walking the major custom AI agent products with an honest eye on their model architecture:
**Microsoft Copilot Studio**: Structurally aligned to OpenAI via the Microsoft + OpenAI partnership. The 'default' model is GPT-class via Microsoft's hosted infrastructure. Other models exist as options but the path of least resistance is OpenAI-via-Microsoft.
**Salesforce Agentforce**: Multi-model on paper. In practice the Atlas Reasoning Engine and the path of least configuration favor Salesforce-controlled infrastructure. Buyers in the Salesforce ecosystem tend to inherit the default rather than pick.
**Writer**: Locked to Writer's own Palmyra LLM family. This is a feature for Writer (vertical integration, single vendor) and a constraint for the buyer (no choice).
**Anthropic Claude Managed Agents**: Claude-only by design. Same constraint shape as Writer, just with Anthropic as the vendor.
**OpenAI ChatGPT Enterprise + Agents**: OpenAI-only by design.
**Google Vertex AI Agent Builder**: Gemini-first, with other models via Vertex Model Garden as configuration overhead.
**Sierra AI**: Publicly claims multi-vendor (OpenAI, Anthropic, Meta) - this is the rare exception among well-known vendors, and Sentie shares the same architectural choice.
**Sentie**: Multi-vendor by default. The substrate routes per-task across OpenAI, Anthropic Claude, Google Gemini, Meta Llama, and open-weight models. Customers do not choose a model; the substrate picks the right one for each request.
The single-vendor pattern is more common than buyers realize. Most agent products implicitly tie you to one vendor's pricing curve, one vendor's outages, and one vendor's strategic decisions.
The Hidden Tax of Single-Vendor Lock
Three real costs that buyers rarely see itemized:
**Pricing concentration.** Frontier model prices are set by the vendor. When OpenAI raises GPT prices (they have, multiple times) or when Anthropic raises Claude prices (same), single-vendor customers absorb the entire increase. Multi-vendor customers route around it - if Claude costs go up, the substrate shifts the workload to GPT or Gemini for tasks where the quality difference does not matter.
**Outage exposure.** Every foundation model vendor has had multi-hour outages in 2025-2026. Single-vendor agents stop working during their vendor's outage. Multi-vendor agents fail over. The cost of an outage is your business stopping; the value of multi-vendor routing is the failover.
**Strategic lock-in.** Foundation model vendors can change terms, deprecate models, restrict use cases, or change capabilities at any time. The agent you built on GPT-4 last year is not the same agent today because the underlying model has shifted. Multi-vendor agents diversify this risk. Single-vendor agents are subject to one company's strategic direction.
None of these costs show up on the agent vendor's invoice. They show up in your bottom line eventually.
How Routing Actually Works
A multi-vendor router is not a coin flip. It is a per-task decision driven by:
**Task type**. Classification, summarization, generation, reasoning, decision-making, voice synthesis. Each maps to a different model tier.
**Stakes**. A low-stakes task (categorize this email) can use a cheap small model with acceptable error rates. A high-stakes task (decide whether to refund $5,000) routes to a frontier model with verification.
**Latency requirements**. Real-time voice responses need a model optimized for low first-token latency. Batch processing can tolerate higher latency for lower cost.
**Context window needs**. Some tasks (analyze a 50-page contract) need a 1M+ token window. Most do not.
**Capability fit**. Some models are stronger at specific patterns (code, math, multilingual, structured outputs). The router knows which model is best at which.
**Cost budget**. The agent has a per-task cost ceiling set by the constitution file. The router picks the cheapest model that meets the quality bar.
**Vendor availability**. If one vendor is rate-limited or down, route around them.
This is not a single algorithm. It is a set of policies the substrate enforces on every request. Sentie's substrate makes these decisions for you; the customer-facing experience is 'the right answer at the right price' without managing model selection.
What to Ask a Vendor
If you are evaluating any custom AI agent vendor, ask these five questions about their model architecture. The answers separate real multi-vendor routing from marketing claims.
**1. Show me the model providers you actually route to today.** A real multi-vendor system has multiple vendors live in production. "We could integrate other models" is not the same as "we route to OpenAI and Anthropic on every request."
**2. What happens when one vendor has an outage?** Real multi-vendor failover is automatic and silent. If the answer involves manual reconfiguration, the routing is not actually live.
**3. How do you decide which model handles which task?** The answer should be specific (task type, stakes, latency, capability, cost). Vague answers ("the best model for the job") indicate hand-waving over architecture.
**4. What happens to my costs if my primary vendor raises prices?** Multi-vendor routers shift workload to cheaper alternatives where quality permits. Single-vendor agents absorb the increase.
**5. Can my engineering team see which model ran which task?** Observability is the test of whether the routing is real. If the vendor cannot show you 'GPT-4 handled this request, Claude Sonnet handled that one,' the architecture is opaque.
At Sentie, all five answers are yes-with-detail. Most agent vendors fail at least three of these tests.
How Sentie Makes the Routing Decision
Sentie's [Business Brain](/blog/what-is-a-business-brain) includes the model router as one of its core components. The decision flow on every request:
1. The agent receives a task (send an email, classify a ticket, decide on a refund, generate a voice response) 2. The router reads the task type from the skill specification 3. The router checks the constitution file for cost and quality policy 4. The router checks vendor health (any outages, rate limits, latency spikes) 5. The router picks the model that best fits task + policy + availability 6. The request runs against the chosen model 7. The action and the model choice are logged to the agent's action log
The customer sees the outcome and the action log. Your dedicated [Success Manager](/about) reviews the routing decisions over time and tunes policy as your business changes. The 'right model for the job' is a property of the substrate, not something you configure per agent.
Why This Matters Now
Foundation model competition is fierce and accelerating. Claude 4.7. GPT-5. Gemini 2.5. Open-weight models from Meta, Mistral, Qwen, DeepSeek catching the frontier on specific tasks. The vendor with the best model for any specific task changes every quarter.
Single-vendor agents lock you to one vendor's roadmap and capability curve. Multi-vendor agents take advantage of the competition. The price of multi-vendor architecture, paid by the agent vendor (Sentie), is the upside of every model improvement, captured by the customer (you).
If you are choosing a custom AI agent vendor in 2026, multi-vendor routing is not a nice-to-have. It is the architectural decision that determines whether your agent's capabilities improve as the foundation model market improves - or whether you wait on one vendor's release cycle for the next gain.
[Start a free Sentie assessment](/onboarding) to see what multi-vendor routing looks like configured to your specific business. Your Success Manager walks through the routing policy during discovery.