Session Overview
Discover how forkable shared logs let AI agents safely query, test, and write to live streaming data without risking production outages.
In this talk, Ram Alagappan, Assistant Professor at the University of Illinois Urbana-Champaign, introduces AgileLog and Bolt. Learn why standard Kafka clusters suffer from performance interference and hallucinated writes when agents explore live data, and how diskless architectures enable zero-metadata-copy branching in just 10 microseconds. Discover how "continuous forks" give agents real-time context with unidirectional isolation, allowing safe validation before changes ever hit the parent log.
What You'll Learn
- The agent challenge: Why speculative probing and LLM hallucinations cause latency spikes and downstream broker crashes
- AgileLog abstraction: Introducing forkable shared logs with severed (S-Forks) and continuous forks (C-Forks)
- Unidirectional isolation: Streaming live parent data to child branches while keeping agent mutations fully sandboxed
- Bolt implementation: Using a diskless storage architecture and lazy tail updates to create forks in 10 microseconds without metadata copy overhead
Chapters
- 00:00 — Introduction & why AI agents need streaming data
- 05:30 — The 4 problems agents create for streaming systems
- 08:30 — AgileLog: Rethinking logs with forking primitives
- 11:12 — Severed forks vs. continuous forks (C-Forks)
- 13:08 — Unidirectional isolation & linearizable ordering
- 16:28 — Building Bolt on a diskless architecture
- 20:14 — Zero-metadata-copy forks & lazy tail updates
- 21:46 — Performance benchmarks & latency isolation
- 24:23 — Summary & open-source research
