- The AI agent development industry has rapidly cycled through four naming stages: prompt, context, loop, and graph engineering, each representing a new layer of scaffolding further from the model.
- Both loop and graph engineering describe patterns of agent reasoning (single-agent cycles vs. multi-agent integration) but are fundamentally reinvented concepts from existing distributed systems.
- The key criticism is that these disciplines focus on the 'shape' of agent reasoning rather than the load-bearing engineering decisions about what the system is built from and how it persists.
- Iii's approach avoids discipline multiplication by unifying agent sessions, traces, and business logic into permanent, reusable artifacts (workers, functions, triggers) on an open-source engine.
- The test for any new agent discipline is whether its artifacts can integrate with the rest of the system permanently, rather than being discarded.