Sunny

← All weeks

What I’ve been reading (W22, 2026)

Staying the author when agents do the typing, why attention is the real bottleneck, and a cloud-infrastructure rabbit hole

  1. Add friction to your agents: write the plan and first draft yourself, use the agent to review, explain, compare approaches, and challenge your thinking, so you stay the author instead of becoming the approver https://vickiboykis.com/2026/05/28/we-should-be-more-tired-than-the-model
  2. Feeling busy is not the same as being productive. You can run 20 agents and feel like a command center, but that does not mean you are shipping 20 agents’ worth of work. Starting agents is cheap. Closing the loop is not. Every output still has to be checked, understood, reconciled, and merged by one person: you. That makes attention the real constraint. Not compute. Not token throughput. Not how many branches are open. Your attention is the scarce serial resource every agent eventually has to pass through. This is an old bottleneck lesson in a new form. Optimizing the non-bottleneck does not increase throughput. It just grows the queue in front of the constraint. https://x.com/addyosmani/status/2059844244907696186
  3. LLMs are most useful when you put them in tool-driven feedback loops that do repeated, verifiable work, rather than expecting one clever prompt to produce genius: think agents grinding through repo migrations file by file, writing tests and rerunning them until they pass, fixing CI failures without getting bored, cleaning messy data against validators, or smashing together prototypes that later harden into real tools. https://x.com/nateberkopec/status/2058014104854626490
  4. Some good principles for designing extremely fault-tolerant systems: isolation, redundancy, static stability. Build from independent parts so no single failure cascades, keep multiple isolated copies of each, and run on the last known good state with enough over provisioning to absorb a dead component’s load. The payoff is decoupling. At PlanetScale a total control-plane failure takes out only admin functions, not customer queries. https://planetscale.com/blog/the-principles-of-extreme-fault-tolerance
  5. A cloud-infrastructure rabbit hole - EC2 (AWS’s cloud server platform) began as a 2003 internal paper to “sell virtual servers as a service” which the core team ended up building from South Africa https://blog.b3k.us/2009/01/25/ec2-origins.html, made me find another thread by Mitchell who thought CloudFormation (AWS’s service for provisioning cloud infrastructure from declarative templates) couldn’t incrementally re-provision infra and shouldn’t be the end-all, called for an open-source, cloud-agnostic alternative, then went and built it himself as Terraform. He’s also the one behind Ghostty, your favourite terminal https://gist.github.com/mitchellh/b52314d30ba22bb76f3d6bb9ff098090