


🎯 Building arrows.to 📻 yagni.fm/episodes 🍊 Karl Pilkington is my spirit animal 🫎 #1 ranked Canadian Football DFS


"Accidentally" clicked in Linear's PR review flow today and it's really good! Very fast and nice touches like grouping changes, the guided review flow, much more pleasant to look at.

Fix your egress with this one simple trick: dont load full ActiveRecord models that have huge json columns just to get two id columns out and then run code a few million times a day

I havent been raving about PlanetScale lately because honestly I havent had to think about our database in a while. Great product, great MCP, highly recommend.
Bring us your Rails apps. I love rails apps.

It feels like there are actually three distinct levels of "run AI 24/7 to look for security issues" 1. Loop to check for published security patches in languages, libraries, platforms you need to upgrade 2. Loop to check your source code for app-specific security issues (pen-test, OWASP type stuff) 3. Loop to red-team your running app (combines 1 and 2) to cover chained issues or production config

We recently rewrote all the READMEs for our repos to: - remove stuff about setting up environments (moved to .agent/setup and targeting a machine reader not human) -add info about production: where it runs, what MCPs can do, project ids in external system, Slack channels to use

This week I added a little feature to delete old export files that we never cleaned up (as they had reached 100GB total) Built the feature then told Amp to report in on the backlog progress in Slack: "post to #talk-code with a summary of what was purged in a table"

"but the agents dont understand the characteristics of our production system" "oh"

So far my approach is: - have an LLM sample a few windows of time and help me write anomaly/threshold/SLO alerts in Cloudflare - ideally in the future, agent can tune/modify those directly - make sure devs have MCP access to logs to help debug/investigate ad-hoc

"just have an LLM scan your logs constantly to look for security issues" -- how exactly are you doing this? Every 5 minutes check the past 6 minutes of logs? If you have 10k logs in that window, you are sending that raw to LLM? Are you having LLM manage platform alerts?

If you've used Heroku CLI in the past, you might have credentials just sitting in your ~/.netrc for any agent to read and you should upgrade the CLI and wipe those heroku.com/blog/securing-her…

Prompt: "we keep getting an email from postmark that our webhook is disabled, check?" 5 minutes later

Yes but...I just say "post lgtm or needs attention" on dependabot PRs Agents know to look at changelogs and check how it's used and run tests etc
Pretty much any dependency upgrade at this point is a single agentic AI prompt. But it’s not just “upgrade from vX to vY. Make no mistakes.” Anecdatally, best results for me have come from something like “Upgrade this codebase carefully from vX to vY, one major version at a time, using upgrade documentation and blog posts about migrations like this one. Ensure all other dependencies keep working with the new versions, and make sure that all tests run green. Do a visual inspection on each upgrade round to ensure functional equivalence. Ask me for clarification if you encounter incompatibility issues.” A bit more wordy, but works. And in the end, still a one-shot event.

tbh I still like PRs as a self-imposed speed bump I can tell agentsto just merge their own for small stuff

I get the sense that most harnesses don't really use the "Resources" part of MCP servers and only "Tools"...is that true? any reading material? Just wondering if there is any reason to put more static stuff in Resources (better aligned to spec)

Back on the mic with @fractaledmind covering why code is cheap but still matters, why software foremans beat factories, and deep breathing exercises for your agents yagni.fm/episodes/i-need-a-s…