The Log Is The Agent - Ishaan Sehgal, Omnara
Summary
Ishaan Sehgal, CEO of Amnara, proposes a fundamental shift in understanding AI agents, arguing that "The Log Is The Agent." He contends that an agent's true identity, history, and state reside in its append-only event log, which records every user input, model output, tool call, and state transition, rather than in the model or runtime environment. This log-centric architecture offers significant system properties, including enhanced reliability, as executors can be fallible and new workers can resume sessions from the log; improved scalability, allowing one process to manage thousands of agents without sticky sessions; natural support for forking agent paths; and simplified migration between AI providers. Furthermore, it enables true multiplayer agent interactions and, critically, ensures user ownership of agent data, preventing vendor lock-in, which Sehgal identifies as a deeper concern than model or API lock-in. Amnara is building an open-source managed agents platform based on this principle, ensuring users fully control their session logs.
Key takeaway
For AI Architects designing robust agent systems, prioritize a log-centric architecture where the agent's durable history is the primary source of truth. This approach ensures reliability, scalability, and seamless migration across providers, mitigating deep vendor lock-in. You should implement append-only event logs for all agent state transitions, enabling easy session resumption, parallel experimentation via forking, and collaborative agent development. Critically, ensure your organization retains full ownership and control over these agent logs to safeguard sensitive data and maintain operational flexibility.
Key insights
An AI agent's true identity, history, and state are its durable, append-only event log, not the model or execution environment.
Principles
- The agent's log is its durable identity.
- All agent operations read from or append to the log.
- Log ownership prevents deep vendor lock-in.
Method
Reconstruct agent state from its log, pass to model for next action, append model response to log. If a tool is requested, run it, append its result, then repeat the loop.
In practice
- Use logs for seamless agent failover.
- Branch logs for parallel model exploration.
- Share log access for collaborative agent work.
Topics
- AI Agents
- Log-Centric Architecture
- Event Logs
- Vendor Lock-in
- Agent Reliability
- Agent Scalability
Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, AI Architect, MLOps Engineer
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by AI Engineer.