ianhxu/agentic-engineering-field-study

This repository is a research report, not a piece of software. It was produced by running several AI research assistants in parallel, each digging through sources published in 2025 and 2026, and then combining their findings into one document. The report tries to map out the fast moving world of using AI agents to write software, a practice it calls agentic engineering, and it comes with eight separate sub reports plus learning resources. The main finding is that this way of working now has a settled vocabulary. Terms like vibe coding, spec driven development, context engineering, the surrounding-setup discipline it calls the environment layer, and loop engineering are not just different names for the same thing. They describe different angles: how much a human stays accountable for the code, whether intent is captured as a durable written spec or through ongoing conversation, and which part of the system engineers currently spend their effort improving. A central idea the report keeps coming back to is that an AI agent's quality equals the underlying model plus everything built around it: the tools it can use, how its input context is managed, the checks that verify its work, and the permissions it operates under. The report cites evidence that changing that surrounding setup around a fixed model produced a large jump in one benchmark score, showing that the system around the model can matter as much as the model itself. The report also argues that practice has shifted from carefully writing single prompts toward building repeating loops: an agent starts with a fresh slate each round, keeps its working notes on disk, and completes one small task per round before repeating. It says giving an agent a way to check its own work is one of the most valuable things a person building these systems can do, since a human reviewing everything by hand becomes the real bottleneck. The report states the core approach it describes can be learned in roughly fifteen to twenty hours: gather the right context, write a plan you can edit, break work into small verifiable tasks, let the agent run its own checks, review with fresh eyes, then commit.