Realtime for AI agents

Agent work outlives a request

A model or agent run may take seconds or hours. It can call tools, produce partial results, wait for a person, or hand work to another agent. The original HTTP request is often over before the user-facing work is done.

HTTP remains useful for starting work and reading from your backend. Realtime infrastructure gives your application a way to show what happens after that request returns: progress, questions, decisions, and completion.

Use State, Channels, and Tasks around the agent you run

Lynta carries realtime information around agent execution. Your application continues to run the agent and own its data.

Keep a person in the loop

A research agent can pause with a question such as “Which market should I prioritize?” A person can choose Enterprise, SMB, or Both; the application records that response and the agent continues. Lynta provides realtime coordination around that flow; it does not execute or steer the agent itself.

The same model can make a coding run, support workflow, background agent, or approval request visible in a dashboard or AI copilot.

Coordinate people and multiple agents

A Task can include Subjects that identify its participants. For example, user:123, agent:researcher, and agent:reviewer can take part in the same live unit of work. Subjects describe participants within a Task; they are not a separate top-level product primitive.

This supports a multiplayer AI pattern: a researcher gathers information, a reviewer checks it, and a person can inspect progress or provide direction in the same application experience.

Your backend remains the source of truth

Lynta does not replace your database or run the AI workload. Your backend decides what is authoritative and runs the agent on the infrastructure and orchestration you choose. Lynta makes the related application changes and events live for connected clients.

Explore the realtime model