I wrote our initial engineering operating model assuming humans would do most of the work.

Over the past month, we started created custom prompts and integrating MCP servers + skills that could read our docs, search our meetings, manage our Asana tasks, review our code, test our implementations, and ship features. And the bottleneck flipped overnight.

The constraint used to be "can we build this?" Now it's "should we ship it?"

Here's what changed: we wired MCP servers and CLI skills into our coding agent (Claude Code ACP which already has nice tools built-in, configured with an open-weight model). Docs, meeting transcripts, Asana, deployment logs, Chrome DevTools. The tools follow the agent regardless of editor (vim, Zed, terminal, doesn't matter). One agent can now plan, implement across multiple repos, self-review against 18 quality checklists, debug end-to-end with browser access, and link everything back to Asana.

Real example: Qwestly Career Agent. ~2 days of human design planning. Agent ships the entire system - backend, frontend, tests, deployment - in hours. QA and paper trail are just part of the loop.

The counterintuitive part: code quality went UP. Those 18 review prompts (architecture, security, accessibility, code & PR review, unit tests, etc.) get applied every single time, not just when someone remembers.

The bottleneck is now the human approval layer. Product & EM signoff gates that were designed for a world where implementation took weeks, not hours. Good problem to have. Still working through how we can improve this bottleneck.

The operating model says what the process should be. The MCP & agent skills ecosystem makes it executable by AI.

Anyone else seeing their process docs outpace the tooling, or the other way around?

#engineering #ai #harnessengineering #mcp #devtools