How I Use AI Agents Day to Day Without Handing Them the Wheel

The boundary is responsibility

I use AI agents most days, but I do not treat the agent as the engineer of record. It can read the repo faster than I can, hold a narrow implementation thread, and put pressure on a diff. The decision, the risk, and the final acceptance stay with me.

That boundary matters because adoption is running ahead of trust. One study of 128,000+ GitHub projects found signs of coding-agent use in 22% to 29% of them. At the same time, the Stack Overflow 2025 AI survey shows a real productivity signal next to very low trust in output accuracy, and METR's randomized trial with experienced open-source developers found AI tools made tasks 19% slower than the control group, even while developers expected and perceived speedups. My takeaway: speed claims need receipts, even when the tool is useful.

The work I delegate

I get the most value when the task has a tight boundary and an easy way to check the result. If the task needs product taste, security judgment, or architectural ownership, the agent can help map the terrain but it does not get the steering wheel.

The loop is smaller than the hype

A normal session is boring on purpose: state the local rule, ask the agent to inspect before changing, keep the diff small, run the command that proves the claim, and review the result as if a teammate opened the PR.

The agent is strongest when it can alternate between discovery and narrow action. 'Find every place this route guard is used' is a good request. 'Modernize auth' is too broad; it invites the agent to guess.

Rules files are part of the workflow

I do not want to re-explain the same constraints in every prompt. Codex can read AGENTS.md, and Claude Code has a memory system built around CLAUDE.md. In an Angular repo, those files should pin the version, the Signals/RxJS boundary, the template syntax, the test runner, and the APIs that are out of scope.

That is why my AI-agent workflow connects to the existing articles on AGENTS.md for modern Angular, CLAUDE.md for modern Angular, and Angular CLI MCP. The standing rules reduce repeated explanation. The MCP layer improves lookup. Neither replaces review.

What I refuse to outsource

I would refuse to merge an agent-produced diff that I cannot explain line by line. That sounds obvious until the diff is large, the tests pass, and the assistant's summary feels plausible. Plausibility is not evidence.

The quiet failure mode is a diff that looks tidy but bends the architecture: a service grows a second state model, a component starts owning data it should request, a test asserts the mock instead of the behavior. The agent can catch some of that in a second review pass. The responsibility still belongs to the developer.