Ryan Scott Brown

I build cloud-based systems for startups and enterprises. My background in operations gives me a unique focus on writing observable, reliable software and automating maintenance work.

I love learning and teaching about Amazon Web Services, automation tools such as Ansible, and the serverless ecosystem. I most often write code in Python, TypeScript, and Rust.

B.S. Applied Networking and Systems Administration, minor in Software Engineering from Rochester Institute of Technology.

TIL: Using Only One-Shot Agents

Models have gotten considerably better in the past year and my intuition for comfortably agent-sized tasks went stale. I now try to one-shot as much as possible by writing better prompts.

I have taken back my editor (it’s human-only now) and I don’t watch inference happen anymore. Agents go out into independent sandboxes (with --yolo) and work without input until it’s time to review and merge.

My new setup is:

  • Zed with all AI features deactivated
  • kitty on Mac or tmux on Linux
  • copilot --autopilot --yolo -p "..." or claude --dangerously-skip-permissions -p "..."
  • Make --yolo safe using colima containers or Lima virtual machines
  • A mounted .env for secrets with appropriate scopes

That’s it, I have stopped using harness-of-harness tools like Conductor/workmux because they’re not better enough to justify how complex they make sandboxing. Limiting myself to a single shot per task has forced me to improve tests, skills, and AGENTS.md directives to keep LLMs on track.

The dopamine hit of juggling fistfuls of agents is a trap. Chat-mode interactions are designed for engagement, not effectiveness. Typing furiously and watching tokens burn is not the same as being productive.

Background and async agents let me plan at my own pace and the work I finish feels more complete. In short, give yourself time to think and build the right thing. My workflow used to be:

  1. Open kitty or conductor
  2. Spin up as many agents as I had tasks
  3. Merge results of those tasks with a mix of automated and manual verification
  4. Continue until I run out of tasks or get tired

Now I’ve decreased the number of open sessions and parallel tasks because evidence shows that humans are terrible multi-taskers and we tell ourselves we aren’t. I try to reduce multitasking by turning notifications and other screens entirely off and having a stack of open Zed buffers to take down ideas and plans.

Turn off any AI tools that interrupt you, quit using interactive chat entirely, and carve a line between human-only and agent-only parts of your work.

Design by Sam Lucidi (samlucidi.com)