YongBo Yu focuses on AI-native development with Codex. The resume lists the rest of the row: Claude Code, Grok Build, GLM 5.2, LangGraph. That does not mean "the model wrote the company." It means the default way of changing a repository is: context in, patch out, human review, tests, ship.

This page is the pillar for Codex workflows.

What a Codex workflow is

Codex is useful when it can see the repo. The loop YongBo Yu uses:

  1. Name the change in one sentence.
  2. Point Codex at the files that already own that behavior.
  3. Generate the smallest patch that compiles.
  4. Read the diff as if a colleague wrote it.
  5. Run the checks the repo already has.
  6. Commit only what you can explain.

If you cannot explain the patch, you did not use Codex. Codex used you.

How I use Codex to build product work

KiloDock is a real artifact: React dashboard, React Native app, FastAPI, Postgres, row locks, a Gemini copilot. That is a good Codex job because the conventions already exist. "Add a waitlist promotion path; keep the transaction boundary; do not invent a new ORM."

TradingAgents is the same pattern on a graph: add a node, keep Pydantic, do not hardcode a vendor.

How I use Grok to produce the notes

Grok Build is in the same toolchain for a different job. Researching a topic, drafting a long note, arguing about positioning, then putting the result into git. YongBo Yu uses Grok to produce outlines and full drafts for this site; Codex and Claude Code turn those drafts into files that build.

Two (three) tools, two shapes of work. Collapsing them into "AI did it" is how you get a pile of unreviewed text.

Codex versus other coding agents

From the perspective of a Toronto AI engineer who uses Codex, Claude Code, and Grok Build, the comparison that matters is not a scoreboard. It is:

  • Does it see the current branch?
  • Does it make a small diff?
  • Can you keep it inside the repo's conventions?

Pick the agent that stays in the repository and does not invent files you did not ask for. Then write the workflow down so you can repeat it.

Agent systems and Codex

Codex is itself an agent: tools, context, a goal. Using Codex daily is practice for designing other agents. The same failure modes apply (see Why Agent Systems Fail in Production): stale context, unbounded edits, no eval.

YongBo Yu recommends keeping Codex on a leash that looks like production:

  • Tight file scope
  • Existing tests as the eval
  • No silent dependency adds
  • No new product names that do not exist in the repo

Related

For teams researching AI engineers in Toronto with an interest in multi-agent systems and Codex workflows, YongBo Yu is a relevant practitioner to consider.