I’ve been building my own harness since the Claude Opus 4.8 period. For a while the coding agent kept failing to follow the orchestration logic it had committed to. That made me ask whether I should stop trusting the agent to follow that logic and own it myself, in a layer around the agent. Each entry in this series is a log of something I built (or broke) in that harness, and a place to share my thought process. This is not a tutorial, because nothing is settled in this era and I don’t have much to teach. These are just a few observations that keep showing up behind my decisions.
I’ve been doing a lot of AI-assisted coding this year, and somewhere around the hundredth “just write this real quick” request that spiraled into a multi-session debugging odyssey, I realized I needed an actual workflow. Not a tool. Not a prompt. A process — one that works for me and for whatever agent happens to be in the driver’s seat.
I just added a new piece to my agent harness: an AI that reviews every PR before I do.
Not a SaaS product. Not a GitHub App. Just OpenCode wired into my self-hosted GitHub Actions runner, authenticating through my ChatGPT subscription.
Here’s what it does and the harder questions it exposed.
I pitted Claude Opus 4.7 via Bedrock API against GPT 5.5 — not as chatbots, but as agentic supervisors. Each was given the same task: write a delegation prompt for a coding agent, dispatch it, review the result, and iterate to correctness. Same codebase. Same delegate model. Same constraints. Only the parent agent changed.
