Xiao Meng

Over the past two years I have kept pace with agentic coding, working each new technique into my daily workflow. A common pattern is to hand everything to the agent, then be disappointed by the result and watch the codebase grow harder to maintain. My experience has been the opposite, I find that agents free me from tedious, repetitive work, so I can put my time into thinking and making decisions.

Writing code by hand is easy to sink into: I could stay in flow for hours and still feel fresh. With agents, I run several projects or features in parallel, and each one keeps pulling me back into the loop to supervise, review, and answer questions. Most of my day now goes to thinking, which is far more demanding than writing code. But judging by the quality of what I ship, I'm convinced this is the right way to work.

Out of this practice came out a complete methodology for building software with agents:

• Spec first: Every feature starts with a planning phase that produces a written spec covering requirements and system architecture. The agent is a sparring partner here, not the author. I use the discussion to refine requirements, weigh architectural trade-offs, and surface gaps and failure modes, iterating until the spec is final.
• Hands-off execution: I break the spec into concrete implementation steps and hand them to a top-tier frontier model. This is the only phase where I deliberately stay out of the way. I review and accept the result afterwards. For complex tasks, I ask the agent to write a session summary along with the git commit.
• BDD as a hard constraint: Agents must implement in behavior-driven style, so every stated requirement is backed by a test. This is where agents pay off most, since they excel at repetitive work: extensive unit tests with mocks, and at times entire mocked services for integration testing.
• Smoke tests as living documentation: Before a major release, the agent smoke-tests the feature and must leave behind a runbook showing how to drive the UI with tools such as Agent Browser or Computer Use. These scripts live in the repo and run alongside the code.
• Context discipline with herdr: I work mostly in the terminal and keep each session's context tightly bounded. When it grows too large, I use herdr, an agent multiplexer, to hand off subtasks or research to fresh agents that report their conclusions back.
• Remote control with paseo: Away from my desk, I connect from my phone to a paseo daemon on my workstation and direct Claude Code or Codex remotely.