Skip to content

agent-engineering.md

Latest commit

 

History

History
76 lines (54 loc) · 2.86 KB

agent-engineering.md

File metadata and controls

76 lines (54 loc) · 2.86 KB

Agent Engineering

Why I use agents in engineering work

I use agents to compress the distance between a fuzzy request and a verified implementation, not to skip engineering judgment.

The value comes from shaping work so it can survive:

  • a bounded brief
  • an explicit execution surface
  • a repeatable verification loop
  • inspectable proof artifacts

If a task cannot be rerun by another engineer with the same repo context and reach the same result, I do not treat it as reliable agent-assisted work.

When I choose Codex vs Claude Code

Codex

  • Use when I need grounded repository work inside the real project.
  • Best fit for source edits, shell-driven inspection, tests, and verification loops.
  • I treat it as the primary execution surface for implementation-heavy tasks.

Claude Code

  • Use when I want an alternate reasoning pass on problem framing, prompt shape, or task decomposition.
  • Best fit for comparing approaches, tightening a brief, or pressure-testing assumptions around execution.
  • I treat it as a complementary lane, not a replacement for repo-grounded verification.

How I make agent work engineerable

Every agent task should be able to inherit the same contract:

Package-manager convention for this repo: use pnpm in briefs, harness commands, and verification steps.

Workflow template

  1. Task brief
    • Goal
    • Changed surfaces
    • Hard constraints
    • Success criteria
  2. Execution surface
    • Relevant repo instructions (AGENTS.md)
    • Prompt or task framing
    • Exact files or subsystems in scope
    • Allowed tools and expected output shape
  3. Verification checklist
    • What command proves the implementation works
    • What UI or route behavior must be visible
    • What docs or metadata must stay aligned
  4. Failure / redo rule
    • If source, docs, and tests disagree, the task is not done
    • If verification fails, rerun the loop instead of reporting a plausible summary

Proof inside this repo

The public proof is intentionally spread across UI, source, tests, and markdown:

  • /tech-stack contains the AI-agent engineering section and the engineerable harness sequence.
  • src/shared/constant/agent-engineering.ts is the canonical source for agent claims, harness steps, and proof links.
  • src/shared/constant/project-detail.tsx includes the react-devtool-cli case study, which is the strongest project-level proof of deterministic agent-facing contracts.
  • src/pages-layer/tech-stack/tech-stack-page.test.tsx and tests/tech-stack.spec.ts verify that the public proof section renders and keeps its evidence links.
  • tests/about.spec.ts verifies that the About narrative also reflects current agent-orchestration focus.

Repo-specific checks

For this portfolio, the main verification path is:

pnpm lint
pnpm test -- --run
pnpm build

That is the bar for calling an agent-assisted change complete here.