The Log Is The Agent
summary
Argues that the agent is the append-only session log not the model or runtime and demonstrates how reconstructing state from that log enables durability, continuity, and forkability. Shows concrete loop: read log, run model, append output, run tools, append results, resume elsewhere; discusses compaction as lossy projection and the ownership risk when a provider controls the log. Practical claims about failover, scaling, and migration are tied to the log primitive rather than vague examples.
AI-written summary · about
chapters (12)
- 0:00Log as the agent
- 1:06Agent is its log not runtime
- 3:32Power of the log from databases
- 4:35Compaction and external state
- 7:04Reliable execution with log
- 7:47Scalability and failover with logs
- 8:29Forking and multiplayer sessions
- 9:30Migration and lock-in avoidance
- 10:51Current agent infrastructure fails
- 11:54Danger of log lock-in
- 12:58Unara builds agents around log
- 14:00Closing and future
related talks
24mLogs as Data
18mEvent-based Design Abstractions for Agent Harnesses
01h 03mAgents Everywhere
01h 00mFrom LLM Theory to Practical Agentic Implementations
12mYour Agent Stack Is Broken at the Bottom
36mAutonomous Policy Validation: Building AI Agents to Analyze Logs and Identify...
26mBeyond Sandboxes: Architecting Durable Runtimes for AI Agents
01h 20mMind the Gap (In your Agent Observability)