AI Infrastructure for Long-Running Agents | Matt Aitken (Trigger.dev)

Let's unpack what it actually takes to build AI infrastructure for long-running agents — reliability, retries, and how fast customers are jumping between frontier models. In this episode of Arena Conversations, Peter sits down with Matt Aitken, founder and CEO of Trigger.dev, to dig into open-source model licensing, computer use, and what Matt thinks developers should build next. Trigger.dev homepage: https://trigger.dev Trigger.dev open source GitHub repository: https://github.com/triggerdotdev/trigger.dev Matt Aitken on X: https://x.com/mattaitken 00:00 Introduction to Matt Aitken 00:23 State of the AI agent infrastructure industry 01:42 Long-running compute as the core problem 02:53 The "Is Claude Up" Slack channel 04:10 Peter’'s agents running for 4 hours straight 07:02 Qwen 3.8 27B's surprising performance 07:38 How fast teams are switching models (June to now) 09:04 Are all agents coding agents? 11:41 Trigger.dev's typical customer use cases 14:17 Iteration speed: deploying every couple of days 16:23 Trigger.dev's origin story and training-data advantage 20:20 Favorite models for infra and optimization work 21:37 Running optimization tasks for 24 hours straight 28:39 What's still missing: context, evals, reliability 29:42 The adoption gap between SF and everywhere else 31:16 Computer use and enterprise adoption 32:31 Which models are best at computer use 33:17 Open source models and licensing 35:28 Closing: check out Trigger.dev's open source repo