Josh Huang's Engineering Notes
Directing an AI Agent Through Production RTL, Not Just Autocomplete
What changes when you point an autonomous coding agent at real chip sign-off work instead of boilerplate — and what 20+ iterations on a production ISP module actually looked like.
Most "AI writes my code" stories are about scaffolding, tests, or glue code — places where being wrong is cheap and instantly visible. Chip RTL sign-off is the opposite: a subtle logic error can survive simulation, survive code review, and only surface as a silicon bug months after tape-out. I wanted to know whether an agentic coding tool (ChipAgents, directed the same way I'd direct a junior engineer) could actually clear that bar on a real production block, not a toy example.
The answer was yes — but only once I stopped treating it like autocomplete and started treating it like delegation to someone who needed explicit checkpoints, not just a prompt.
The block: ISP Auto Focus, end to end
The target was a full Auto Focus module for an image-signal-processing subsystem — not a stub, the actual production block: RTL, C-model co-simulation for bit-true validation, coding-style compliance, and timing closure. This is the same bar any human-authored deliverable in this subsystem has to clear before it ships.
Unlock this article to keep reading, or subscribe for unlimited access to everything. See Pricing for details.