Summary
Dan Farrelly, CTO and co-founder of Inngest, takes the AI Engineer conference stage to argue that every agent system has a measurable “half-life” — and that most teams are building in ways that guarantee expensive rewrites every six months. His central framework divides agent architecture into three discrete layers with very different decay rates: prompts (weeks), models and context (months), and execution infrastructure (years, if designed correctly). The problem, he argues, is that teams routinely couple all three together, causing the fastest-decaying layer to drag down the rest.
Farrelly defines the execution layer as the system responsible for running code reliably — managing flow, state, durability, retries, and orchestration — independently of the infrastructure it runs on. This is the layer worth investing in carefully, because it is the only one stable enough to outlast multiple model generations and framework rewrites. He identifies several common anti-patterns: using sandboxes for state management (sandboxes are ephemeral by design and should receive context from the execution layer, not store it), burying orchestration logic inside frameworks where it can’t be swapped, and conflating prompt logic with retry and backoff handling.
The talk closes with a forward-looking review of emerging agent patterns — background agents, autonomous loops, agent factories, dynamic workflows — and argues they share a common requirement: full session-level observability from trigger through every tool call, database error, and permission failure. Farrelly draws on Inngest’s production experience building execution infrastructure for long-running asynchronous agents to ground each recommendation in real engineering constraints.
📺 Source: AI Engineer · Published July 21, 2026
🏷️ Format: Deep Dive
