Define the agent boundary
An agent request needs instructions, relevant context and a clear set of permitted actions. Keep these inputs explicit in your application rather than expecting a raw model call to inherit workspace configuration.
Start with a read-only task such as answering a question from retrieved documents. This lets you evaluate source use before introducing tools that change external systems.
Connect retrieval to the model request
Retrieve authorized passages, attach stable source identifiers and include them with the user’s question. Keep the context within the limits of the selected model and your latency budget.
Send the request through TextCortex using the available model identifier. Use the same integration to compare proprietary and open-weight candidates on answer quality.
Execute tools through application controls
For a model route that supports tools, validate the proposed tool name and arguments against your application’s rules. Enforce permissions before execution and return the result to the model only when appropriate.
Use bounded steps and timeouts. Require review for consequential actions, and prevent a repeated model request from executing the same external action twice.
Route by agent step
A classification step and a complex synthesis step can use different models. Keep the mapping configurable and record which model produced each result.
Use the routing guide to organize this mapping. For EU workloads, constrain every agent step to the permitted deployments, including fallback paths.