Most “responsible AI” content reads like it was written by a policy team that has never deployed an agent to production. The checklists are long. The principles are abstract. And none of them tell you what to do when your agent starts hallucinating customer data at 3 AM and the on-call engineer is asleep.

I have been building AI agents for about a year now. Not research. Not demos. Actual agents that touch real data, make real decisions, and occasionally break things in ways I did not anticipate. Here is what responsible AI looks like from the builder’s side — not the policy side.

—

Read more

Lately I have been feeling like twenty-four hours is not enough. Not because I have too much work, but because every single hour seems to bring something new I want to learn. A new framework. A new tool. A new piece of research. A new opinion from someone I respect. By the time I close one tab, three more have opened in my head. I catch myself doing the math in the shower — sleep seven, work eight, pray and eat and commute, that leaves maybe three or four hours of focused time. Not a lot when the input keeps multiplying.

GROW coaching framework

And yet most people I know are not running out of time. They are running out of attention. They have all the time in the world, and they spend it on things that, if we are being honest, do not move the needle. Endless scrolling on TikTok. Reels until the battery dies. Mobile games that reset every morning. I am not pointing at anyone else here. I have been that person. Sometimes I still am.

So the question I keep coming back to is simple: how do we handle so much information in an era that keeps producing more of it than any one human can absorb, and how do we make sure the time we do have is spent on things that actually matter?

Read more

Last year a client asked for an AI agent to automate their customer complaint triage. Forty hours of scoping done, two weeks of build time blocked. I was three days from opening the IDE when I sat next to the customer service team for two hours and watched.

I expected to see overwhelmed agents drowning in tickets. What I saw was three CS staff handling 12 complaints a day each, perfectly fine, with one exception — they refused to escalate anything to the operations team. Not because escalations were hard. Because the SOP for escalation was 11 steps long, contradicted itself in steps 4 and 7, and the last person who escalated “incorrectly” got a written warning.

The complaint volume was not the problem. The fear of escalation was. The triage agent I was about to build would have processed tickets faster into a system the team was already afraid to use.

I killed the project. Rewrote the escalation SOP from 11 steps to 4. Ran a 30-minute training. Two weeks later, complaints were down 60%. Zero lines of code shipped. The client saved 40 hours of dev time and $8K of cloud budget.

Since then I have had the same conversation five more times. Different industry, different “we need an app” request, same ending: a procedure nobody had bothered to write down clearly was doing the work that 200 hours of engineering could not.

This is the part of design thinking nobody talks about. Most of the time, the right output of the empathy stage is not a prototype. It is a one-page procedure that everyone can read, agree on, and follow. The build comes later — if it comes at all.

Read more

Last week I read a paper that made me re-evaluate everything I have written about AI agent optimization. Microsoft and three Chinese universities published a method called SkillOpt. The result: a single Markdown file, between 300 and 2,000 tokens, lifted GPT-5.5 by an average of 23 points across six procedural benchmarks. No fine-tuning. No new model. No extra tools. Just a Markdown file that gets fed to the agent as context at inference time.

The skill beats handwritten instructions, one-shot LLM-generated instructions, and four specialized training methods (Trace2Skill, TextGrad, GEPA, EvoSkill). It works in Codex. It works in Claude Code. It transfers across model sizes. A spreadsheet skill trained in the Codex loop lifts Claude Code to the same level as a skill trained directly in Claude Code.

After reading the paper, I stopped adding features to my AI agent for a week. I started writing skill files instead. Five of them. All under 1,000 tokens. All producing measurable improvements in my daily work. This post is the five skills and the pattern I now use by default for any procedural task.

Read more