The same model can behave very differently under a different harness. This article examines model ergonomics: the fit between models, tools, context, feedback, and execution environments, using Grok on the Codex harness to explore the boundary between generalized and specialized harnesses.
LLM agents should not be understood as a chain of tool calls; as tasks grow longer and environments become more complex, the runtime needs an event-driven architecture for state, concurrency, observation, and recovery.
Instead of understanding LLM agents as models calling tools, we can view them as systems made up of communicating processes and coroutines; this perspective better explains event sources, control transfer, suspension, and recovery.