williamliu.ai
EN中文
← Blog

The Real AI Coding Race: Which Tool Best Encodes Engineering Judgment

AI coding tools compete on how well they encode engineering judgment, including evidence gathering, verification, boundaries, and review, rather than agent count.

3 min read
  • AI coding
  • Engineering judgment

TL;DR

  • AI-native engineers already carry a personal operating system; tools should make it explicit.

  • Dynamic workflows are promising, but more agents is not automatically better engineering.

  • Agent swarms can produce expensive nonsense: token cost, synthetic consensus, hidden complexity.

  • The real question is bigger than Claude Code — it applies to Codex, OpenCode, and any agent.

  • The winning tool will expose better ways to encode judgment, not just a stronger model.

🧭 AI-native engineers already have workflows

Every AI-native engineer I know has a personal operating system. It is not always written down, but it exists.

After 20+ years of engineering, I do not just "ask for code." I bring a workflow: read the existing system before changing it, find the smallest safe change, define what "done" means before implementation starts, isolate risky edits, make verification explicit, review the diff like it will wake me up at 2am, and leave context for the next run.

AI agents do not remove that process. They make its quality matter more. With a weak workflow, the agent simply moves faster in the wrong direction. This is the argument I made in Encoded Judgment Becomes the New Moat: once that judgment is structured and reusable, it stops being trapped in individuals and becomes infrastructure — and that was always the real point of treating agents as more than "just a tool".

⚠️ The risk: agent swarms can manufacture expensive nonsense

I am optimistic, not blind.

Anthropic notes that dynamic workflows can use substantially more tokens than a typical Claude Code session, and the docs describe limits on concurrency and total agents. That is good product hygiene — but it also tells you this is not a free abstraction. Token cost is not theoretical: I recently traced a single background agent that quietly burned over a billion tokens against my weekly quota.

More agents can also mean more duplicated reasoning, more surface area for wrong assumptions, more false confidence from synthetic consensus, and more hidden complexity behind a polished final answer. A polished, high-scoring output is not the same as a correct one — I've argued before that a high score isn't a good score.

So dynamic workflows should not become the default mode for every task. They are a tool for work where process quality changes the outcome.

🛠️ This is bigger than Claude Code

The better question is not "is Claude Code's feature useful?" It is: what tools help engineers use AI coding agents better — across Claude Code, Codex, OpenCode, and any other serious agentic environment?

Models will keep improving. Teams still need control surfaces around them: the task contract, what evidence to gather first, which boundaries are off-limits, when to ask for approval, what verification must run before completion, when another agent should review the result, and what to save for future context.

Those are engineering-process questions, not model-specific ones. The winning tools will not just expose stronger models — they will expose better ways to encode engineering judgment.

This is the second of three pieces. Next: I put dynamic workflows to work on real projects and report what survives contact.