Fork the Stream: Safe, Isolated AI Agents on Real-Time Data with AgileLog

AI agents are starting to interact directly with streaming data: querying live topics, generating test events, building stream processors, and even writing records back into real-time pipelines. That is exciting, but it is also dangerous. Agents are exploratory, expensive, and sometimes wrong. If they run against the same production streams as regular applications, they can create latency spikes, interfere with critical workloads, or publish bad records that downstream consumers immediately act on.

This talk introduces AgileLog, a forkable shared-log abstraction for making agentic workloads safer on streaming systems. The core idea is simple: give each agent an isolated fork of the stream. Unlike a point-in-time snapshot, AgileLog’s continuous forks keep receiving new records from the parent stream while keeping the agent’s own writes private. Agents can analyze fresh data, inject synthetic events, test stream processors, validate proposed writes, and promote only safe results back to the main log.

We’ll cover the design of continuous forks, severed forks, squash, and promote; explain how Bolt implements these ideas efficiently using a diskless log architecture; and walk through real agentic streaming examples including IoT analytics, stream-processor testing, and supply-chain automation.

Attendees will leave with a practical framework for thinking about safe AI agents on streaming data, why existing log abstractions need to evolve, and what future streaming platforms may need to support as agentic applications move into production.