opinions my own

Simple Agent Patterns

My practices for agentic engineering

Code Got Cheap. Decisions Didn't.

Coding agents dramatically drop the cost of typing code into the computer, which disrupts so many of our existing intuitions about which trade-offs make sense.
— Simon Willison, Agentic Engineering

Code is cheap, but good decisions are expensive.
Use multiple agents in sandboxes
Give the agents tools to show their work
Don't make slop

So, you may already be reading Simon Willison - you should - who has been writing recently about how the old patterns of work don't really fit anymore. The world hasn't caught up to the fact that code is basically free now. And that the way we do work doesn't reflect that change, and we'll have to come up with new ways of doing work, and no one knows what that looks like. This is still so new and fresh that we don't have the tools to figure it out. But we will figure it out. And here are some guidelines I use to make sense of the world. My workflow is not fancy in any way. I have a few skills. I think the CLIs and agents today are really quite good, and you can get a lot done. The agent harness isn't even that important. Copilot is good. Claude is good. Opencode is good. They are all really good. Pick one or use them all. Try all the models. And try them in parallel. And make it easy to do parallel work. Let's discuss

Prompts for Thinking Clearly

But the agents can help us make better decisions:

/explore — learn a new codebase quickly
/plan — always plan your changes, and share the spec
Try all the options before deciding. Then iterate to make the code "good"

"Good" = correct, simple, maintainable

Parallel Agents

More iterations → faster data → better decisions

cd ~/projects/my-app

# each agent on its own worktree and branch
git worktree add ../my-app-A -b agentA/feat/add-search main
git worktree add ../my-app-B -b agentB/feat/add-search main

cd ../my-app-A && \
  copilot -p "implement search as described in @SPEC-A.md, create a PR" &

cd ../my-app-B && \
  claude "implement search as described in @SPEC-B.md, create a PR" &

Parallel Agents → choose your sandbox

Codespaces

Scoped GITHUB_TOKEN, pre-installed tools

gh codespace create \
  -b agentA/feat/add-search

> copilot -i \
  "implement the search bar feature"

CCA

Firewalled, no org secrets, workflow approvals

echo "Fix all failing lint.
Do not change business logic.
$(gh pr checks 75)" \
  | gh agent-task create -F - --follow

⚠ Think about the task and what the agent has access to.
Secrets + unrestricted network + arbitrary instructions = 😬

Parallel Agents: Let them demo their PR

Make reviews easier with a preview

Preview tools: vhs, agent-browser, rodney, playwright, make your own!
Agent explores → previews all approaches → commit to one and make it good

🎁 Bonus: agents read a11y elements. If the agent can't navigate, neither can a screen reader user — fix that.

Don't Make Slop

Be the first to review AI-generated code.

/review — and try different models
Comment on your own PRs
Check and resolve CCR comments
Include the prompt (/share)

Review your code — Review your own PR first

Show your work — Share the prompt and plan

Recap

Code got cheaper, and decisions stayed expensive.
Agents can make decisions easier, but working in parallel sandboxes…
…and letting them show off their work…
…and you should show your work too.

This approach will intensify existing friction in security, accessibility, CI times, preview deploys, and code reviews.