Uploaded September 2026 | Updated November 2026, 13 hours ago
AI agents can monitor live data, investigate anomalies, test hypotheses, and take actions. But today's streaming systems were built for predictable programs, not autonomous agents that are dynamic, exploratory, and sometimes wrong. This creates two fundamental problems. First, agentic workloads can interfere with production workloads. Second, letting agents write directly to live streams is risky: a bad decision can result in catastrophic failure of downstream systems.
In this talk, Ram Alagapan argues that the shared log, the core abstraction behind modern streaming systems, should become forkable to support agents acting on live data. A fork gives an agent an isolated instance of a live stream in which it can explore, run what-if analyses, and stage changes without disrupting production.
He presents AgileLog, a forkable shared-log abstraction with novel primitives designed for agents operating on live, streaming data. And Bolt, an implementation of AgileLog, demonstrating how it makes forks very cheap while providing logical and performance isolation.
00:35 Introduction
02:15 Agents are Forcing Us to Rethink The Log
03:30 AI Agents are The New Users of Data Systems
09:48 The Conventional Log Abstraction Underpinning Stream is a Mismatch for Agents
15:54 The Log Must Become Forkable
25:30 Making Forks Practical
31:34 Bolt's Approach
37:17 Forkable Logs Are More General
39:20 Summary
Talk from Systems Distributed '26: systemsdistributed.com
linkedin.com/in/ramnatthan-alagappan-09953935
AI agents can monitor live data, investigate anomalies, test hypotheses, and take actions. But today's streaming systems were built for predictable programs, not autonomous agents that are dynamic, exploratory, and sometimes wrong. This creates two fundamental problems. First, agentic workloads can interfere with production workloads. Second, letting agents write directly to live streams is risky: a bad decision can result in catastrophic failure of downstream systems.
In this talk, Ram Alagapan argues that the shared log, the core abstraction behind modern streaming systems, should become forkable to support agents acting on live data. A fork gives an agent an isolated instance of a live stream in which it can explore, run what-if analyses, and stage changes without disrupting production.
He presents AgileLog, a forkable shared-log abstraction with novel primitives designed for agents operating on live, streaming data. And Bolt, an implementation of AgileLog, demonstrating how it makes forks very cheap while providing logical and performance isolation.
00:35 Introduction
02:15 Agents are Forcing Us to Rethink The Log
03:30 AI Agents are The New Users of Data Systems
09:48 The Conventional Log Abstraction Underpinning Stream is a Mismatch for Agents
15:54 The Log Must Become Forkable
25:30 Making Forks Practical
31:34 Bolt's Approach
37:17 Forkable Logs Are More General
39:20 Summary
Talk from Systems Distributed '26: systemsdistributed.com
linkedin.com/in/ramnatthan-alagappan-09953935










