Decouple Agent Layers or Rewrite Every 6 Months

Most teams tightly couple prompts, models, and orchestration inside agents. Prompts change weekly and models monthly, so the system decays every few months. Dan Farrelly proposes a three-layer model: context (models, prompts), compute (sandboxes), and execution (state, retries, orchestration). A durable execution layer can outlast years of change in other layers.

Without this separation, a new model or architecture forces a rewrite of orchestration, tools, and retries. Background agents, autonomous loops, and agent factories demand long-running, inspectable workflows that break conventional designs. A dedicated execution layer provides the observability and coordination hub these patterns need.

The argument is timely but vendor-shaped; the execution-layer concept maps to the speaker’s product. Yet the principle—design an orchestration kernel to outlive transient AI surface—is a useful north star, even with existing durable primitives like queues or workflow engines.

Treating sandboxes as ephemeral hands, never as home for state, is a key operational takeaway. The larger tradeoff: accepting an additional infrastructure abstraction that you must now trust not to become the next bottleneck. The talk lacks production case studies and relies heavily on plausible reasoning rather than empirical data.