What off-the-shelf got right
Claude Code, Codex, and Cursor cover a lot of ground. For coding loops inside a repo, they are hard to beat. Most teams should start there.
Where it stopped
Per-client isolation. Credential scope. Receipt schemas that match our Notion workspace. Tool contracts that survive model swaps. Off-the-shelf harnesses optimize for the single-developer case.
What we built
A thin harness that runs each client in a sandbox with their own secrets in Infisical, offloads credential-bearing tools to Cloudflare remote MCPs, and writes a receipt row to Notion for every tool call.
When we still reach for a vendor harness
For internal coding work, we use Claude Code. For browser-driven tasks, Perplexity Computer. The custom harness is for client-facing agent systems where we own the surface.