Subscribe
Sign in
Home
Notes
Start here
Loop Lab
Podcast
Newsletter
Archive
About
Latest
Top
Discussions
Can an AI decision model stand in for your manual tester?
CI can’t tell you whether a person can actually use your app. I tested whether an AI player can cover that click-through, for about 29¢ a game.
Nov 4
•
Ryan E. Hamilton
2
2
October 2026
Agents will find your docs. Have you checked if they're still right?
Here’s what happened when I tested our onboarding flow and found that agents have multiplied the paths into our product (and the docs that need…
Oct 30
•
Rosie Yohannan
2
Your AI software factory has an inventory problem
Why work in progress is becoming the next big challenge for AI-assisted development.
Oct 22
•
Confident Commit
3
We stopped doing migrations by hand. A bot does them now.
How a scheduled CircleCI pipeline with LLM-powered fixers replaced hundreds of manual pull requests… and keeps running forever.
Oct 21
•
Claire Alvis
2
When MCP is present, it always wins
An experiment reveals that agents ignore explicit CLI instructions when MCP is available. The fix? Prompt engineering, not tool configuration.
Oct 13
•
Joaquin Sandoval
1
1
Benchmarking System One models (Jev, Laya) before trusting them in CI/CD
A small elevator simulator benchmark shows hosted and open-weight models still struggle to beat a simple rule zero-shot.
Oct 8
•
Fabio Ramirez
1
External memory cut our agent's flip rate by a third. Cost: 7% more tokens.
Adding a note-taking tool to a pipeline-fixing agent cut its inconsistent "flip rate" by a third, with no drop in fix rate or merge quality and only a…
Oct 6
•
Joaquin Sandoval
1
I installed WordPress on a CI microVM. It turned into a plugin dev environment.
What started as a curiosity experiment turned into a surprisingly practical workflow for testing WordPress plugins across multiple PHP versions…
Oct 1
•
Okamoto Hidetaka
1
1
September 2026
I cut CI failure context 40x using the --failure-report flag. Retries got cheaper, not fewer.
Swapping raw CI logs for --failure-report cut failure context 40x with no change in retry count: cheaper to read, not fewer times reading it.
Sep 29
•
Ryan E. Hamilton
3
Schrödinger's prompt: Truncated or not, depending on when macOS was looking
macOS silently drops deep-link prompt content past 1,024 bytes, and whether your prompt survives depends on what the shell happens to be doing at that…
Sep 24
•
Rodrigo Palomares
1
1
What we've learned (so far) making CircleCI work for agents
Five things we learned about context, tools, and evaluation when we started building for AI agents instead of humans
Sep 23
•
Confident Commit
3
Bake low Merge Efficiency Ratio (MER) into the agent loop
How engineering your agent loop for low MER (with inner-loop checks, a CI sneak peek, and a single push) cuts outer-loop CI credits by 5–6x without…
Sep 22
•
Ryan E. Hamilton
2
1
This site requires JavaScript to run correctly. Please
turn on JavaScript
or unblock scripts
