Is the moat the model or the loop? | First Pass Ep. 2 with Dan Farrelly (Inngest)

Summary

This episode reframes AI agents as production loop systems rather than prompt-centric demos. Dan Farrelly of Inngest argues that the important abstraction is the loop: a continuously running decision layer that inspects state and decides what to do next, backed by skills (workflows, tasks, or sub-agents) and an orchestrator that manages invocation, completion, timing, and reliability. The hosts connect this to Satya Nadella’s idea that the moat is the loop, with feedback and usage data compounding over time. The practical takeaway is that builders should focus less on model selection and more on orchestration, observability, failure recovery, and the ability to improve systems over time.

Notes

• The episode frames AI agent loops as the next abstraction beyond one-shot prompting.
• The hosts say prompting is giving way to loop-based agent design and iterative goal functions.
• Farrelly contrasts early LLM one-shots with higher-level systems that supervise repeated actions.
• He defines loops as higher-level goal-setting for systems, including agents and tasks.
• A production loop must run continuously against system state, not just in a local coding environment.
• Farrelly splits the architecture into three layers: loop, skill, and orchestrator.
• The orchestrator handles invocation, completion checks, scheduling, and durable 24/7 operation.
• He argues the real problem is the system that runs the model, not the model itself.
• The architecture must account for triggers, timing, state evaluation, skill invocation, and failure recovery.
• The hosts compare loops to autonomous workflows with state, retries, and failovers.
• Loop systems are harder because they operate at a higher level and manage many tasks.
• Farrelly says the end goal is often to replace a human in the loop for operational triage.
• Observability is required so the loop can tell whether a task failed and whether a skill was effective.
• Farrelly ties Nadella's loop idea to human capital, token capital, feedback signals, and compounding.