LLM agents in the engineering loop

GenAI is only interesting to me once it is load-bearing. The point is not a demo that impresses in a meeting. It is tooling that quietly removes work from senior engineers every day.

The idea

Most engineering toil is judgment-light and context-heavy: triaging a failing pipeline, keeping docs current, doing the first pass on a code review. That is exactly the shape LLMs are good at, if you wire them into the systems where the work actually lives, not into a chat window off to the side.

How it works

• Agents on the rails that already exist. MCP-integrated agents plug into CI/CD and observability, so an agent triaging a build sees the same logs, metrics, and history an on-call engineer would, and acts in the same place.
• Automate the judgment-light layer. DevOps triage, documentation, and first-pass code review run through agents, freeing senior engineers for product work instead of the work around it.
• Multi-LLM, not single-vendor. For reviews and plans I dispatch Claude, Codex, and Gemini in parallel and synthesize the results. Orchestration stays on the strongest model; deterministic work gets pushed to cheaper or local ones to keep token spend honest.

What it unlocked

Reviews get broader coverage and fewer misses, because three models with different blind spots rarely miss the same thing. Triage and docs stop being a tax on the people best able to do hard work. And it reinforces a view I hold strongly: treat AI as useful, not magic. Wire it into the loop, measure what it actually removes, and do not oversell it.