Dan Farrelly

Dan Farrelly is co-founder and chief technology officer of Inngest, where he builds durable execution for AI agents and workflows: infrastructure that preserves progress, retries failures, and coordinates software running across unpredictable tools, models, and machines.

Farrelly began his career in civil engineering and construction management before teaching himself web development. After working at the independent-film startup Slated, he joined Buffer as a front-end engineer in 2014. He moved through full-stack engineering, infrastructure, and operations before becoming Buffer’s chief technology officer. His independent projects include MailDev, an open-source email-testing tool, and Timezone.io, a utility for distributed teams.

At Buffer, Farrelly encountered the practical difficulties of processing webhook traffic that arrived in bursts and depended on unreliable downstream services. Those experiences informed his approach to scalable event-driven systems. In 2021, he co-founded Inngest with Tony Holdstock-Brown; the company released its workflow platform in 2022. Farrelly subsequently introduced self-hosting alongside Inngest’s open-source 1.0 release, allowing developers to run durable workflows on their own infrastructure.

• Separate execution, context, and compute. Farrelly divides agent architecture into durable execution; rapidly changing models, prompts, tools, and memory; and compute environments such as sandboxes and browsers. Keeping these layers independent lets teams replace models or runtimes without rebuilding the system that manages state, retries, scheduling, and coordination. His agent-architecture analysis treats execution as the layer most likely to survive successive technical changes.

• Build a durable agent harness. An agent harness should persist progress outside ephemeral sandboxes, retry failed steps without restarting completed work, and support events, scheduled jobs, delegated agents, and human intervention. Full-session observability should include database failures, permissions, triggers, and infrastructure—not only model calls.

• Connect agent loops to measurable outcomes. Farrelly describes operational agent loops in which scheduled health checks trigger triage agents and reviewers inspect execution history to improve subsequent runs. His approach to outcome-based agent evaluation connects those traces to concrete consequences: whether engineers acted, a pull request was opened, or a research report was saved. That execution-aware feedback makes asynchronous agents inspectable and measurable after they leave the chat interface.