Really nice deep dive here: this is exactly the kind of “architecture X‑ray” teams are going to need as coding agents move from toys to real production work.
As an independent builder working on agentic and MCP-heavy systems, I really appreciate how ArchSteer turns a big, star-heavy repo into a concrete, inspectable architecture dataset instead of another vague diagram. The focus on hotspots around agent.py, run state, and tool execution feels spot-on; those are the exact seams where agents can quietly bend an architecture while tests still pass.
The “map -> intent -> steer -> check” loop and the ratchet on net-new drift are especially encouraging - that’s the kind of pragmatic control layer that makes it realistic to let agents touch core orchestration without losing the plot over time. Love seeing this level of rigor applied to the OpenAI Agents SDK, and it definitely makes me more excited about pairing ArchSteer with our own agent frameworks.
Keep watching archsteer.com, architecture concern is one of missing piece worries me agentic ai era. Keep up the good work.