Agents name a route. The route names the model.
AEGIS Intelligence is designed to keep your agents independent of any one vendor. Models will keep changing, and agent definitions should not have to change with them.
Route: planning-default
Simulated- PrimaryReasoning model ACommercial API
- Fallback 1Reasoning model BAnother provider
- Fallback 2Open modelSelf-hosted endpoint
Routed to Reasoning model A
Fallbacks and stops are decided in advance.
Take providers offline and watch the route respond. If nothing is left, or the budget is used up, the step stops with a stable error code instead of an improvised answer.
A fallback may behave differently from the primary, so it should be allowed only where evaluations show it is acceptable.
Key capabilities
What the layer handles.
Provider adapters
Native APIs and one adapter for OpenAI-compatible endpoints, using your own keys in the first release.
Model routes
A named primary with ordered fallbacks and the error types that trigger them.
Token and cost tracking
Usage per step, agent and project, priced from a table you can inspect.
Budgets
Limits per run and per project. Reaching one stops the work cleanly.
Policy-based routing
Choosing by cost, capability or data residency, with each choice explainable. Planned for Phase 3.
Neutrality test
Neutrality is a test, not a slogan.
Our release criterion is that the same reference plan completes on at least three providers from identical configuration. Behavioral differences are recorded, not hidden.
Add a provider
Enter an endpoint and your key. It is stored like any other credential.
Create a route
Choose a primary and fallbacks, and set a budget.
Test
Send a sample request through each model in the route.
Use it
Point an agent version at the route.
Change later
Swap models without editing the agent, and compare results through evaluations.
NVIDIA ecosystem
Optional adapters, each with its evidence.
We are evaluating NVIDIA NIM microservices and Nemotron open models as selectable routes for planning and tool-calling steps. NIM exposes a familiar API shape, so it would use the OpenAI-compatible adapter. No such integration exists today.
These would be optional, and AEGIS would work the same without them. Production use of NIM requires the customer's own NVIDIA AI Enterprise license; FLEXVOTIX does not resell it.
NVIDIA, NIM, Nemotron, NeMo and Dynamo-Triton are trademarks of NVIDIA Corporation. They are named here to describe planned technical compatibility. FLEXVOTIX has no partnership, certification or endorsement from NVIDIA.
See the evidence ladderRelated capabilities
Tied to one model provider today?
We would like to understand what makes switching difficult for your team.