Hosted inference API

Rather than running a model on its own infrastructure, a Beam agent can send a piece of text, an image, or another input to a hosted model through a single API call and get a prediction back. This suits tasks like classifying a message or scoring sentiment on a customer note. The agent applies the approved next step based on that output, such as tagging a record or choosing a routing path. Where the model returns a low confidence score or an unexpected result, the agent routes the case to a person instead of acting on it.

Model Hub catalog

The Hub holds a large public catalog of pre-trained models covering text, images, audio, and other tasks, each with documentation on what it was trained to do. An account can pick a specific model suited to a task, such as one tuned for sentiment or one tuned for a particular language, rather than using a generic default. A Beam agent calls whichever model the account has approved for that step in the workflow. If a task falls outside what the selected model reliably handles, the agent routes the item to a person rather than forcing an answer from the wrong model.

Custom inference endpoints

Beyond the shared public models, an account can deploy its own model, including a fine tuned or private one, on a dedicated endpoint built for production traffic. A Beam agent calls this endpoint the same way it would call any hosted model, sending an input and receiving a result tied to that specific deployment. This matters when a task needs a model trained on the account's own data rather than a general purpose one. Failed calls, timeouts, or results outside the expected range are routed to a person to check the endpoint or the input before retrying.