Microsoft Foundry and Twilio: One Agent, Every Channel

Microsoft and Twilio are stitching together a single AI agent that can move between a phone call and a text message without forgetting what the customer just said. The architecture uses Microsoft Foundry for the model layer and the agent runtime, Twilio for the voice, SMS, and chat channels, and an open source SDK called Twilio Agent Connect to bridge them.

The problem most conversational agents share is amnesia. A customer tells a text bot that their flight is delayed, then calls the same company an hour later and the voice bot starts from zero. That gap is not a model limitation. The models themselves are strong and cheap to reach. The gap sits in the layers around the model: the channels, the memory that should follow a customer across those channels, and the runtime that has to keep everything reliable under real traffic.

Microsoft Foundry supplies the models and a new kind of compute built specifically for agents. The runtime is designed to hold state across turns, across channels, and across sessions. Twilio Agent Connect is the SDK that lets that runtime talk to Twilio's phone, SMS, and chat endpoints without custom middleware for each one. The result is a single agent that sees a voice call and a text message as the same conversation with the same memory.

The architecture works like this. A customer calls in. Twilio routes the audio to the agent runtime running on Foundry. The runtime transcribes the speech, passes it to the model, gets a response, and sends it back as speech or text depending on the channel. If the customer switches to SMS mid-conversation, the runtime picks up the same session ID and the same context. The model does not need to be retrained or fine-tuned for each channel. The channel logic lives in the SDK, not in the model weights.

Deployment is a single command. The SDK and the runtime are packaged together so that a developer can spin up the full stack – Foundry model endpoint, Twilio channel integration, agent memory – without wiring each piece by hand. Microsoft and Twilio are betting that the bottleneck to multi-channel agents is not the AI but the plumbing. The SDK is the bet on that plumbing.

For companies already using Twilio for customer communication and Foundry for AI workloads, the integration removes the custom engineering that usually kills multi-channel projects. The agent runtime handles the state. The SDK handles the channel translation. The model just answers questions. The amnesia problem, at least in theory, is solved at the architecture level rather than the prompt level.