The feed

OpenAI models are very good, especially Astra, but they have significant limitations. They frequently ignore AGENTS.md instructions as well as explicit instructions in loops and chats. They are infuriatingly sloppy at cleaning up unused worktrees + files, and with multiple concurrent workstreams, can gobble up resources in a moment.

In some ways, they are more autonomous than ever and I can let them crank on extremely long running tasks. In most practical ways, I still have to watch them closely to make sure they don’t go off the rails or tank my working machine. And I have to budget a decent amount of my attention and tokens to clean up after them.

The current state of working with AI feels a whole lot like a plate spinning act. Even the loading wheels on session tabs in OpenAI’s and Anthropic’s desktop apps remind me of plates going round and round …

If I’m being more accurate, it feels like multi-dimensional looming. That more accurately resembles the felt sense of weaving something out of nothing on multiple concurrent sequences that ultimately connect at some point in the future.

But plate spinning is more viscerally intuitive, and slightly comical. So I’m going to stick with that for now.

Token debt is having old recurring loops or cron jobs that are still running but are effectively useless. Finding I have a bunch of those floating around.

One of the skills of working with AI is parallelizing work between sessions, models, and harnesses efficiently. Another is knowing how to stage long-horizon agent tasks to run in the background while doing the mandatory human review tasks in the foreground.

It’s really easy to forget how deep in the AI echo chamber I am. Being at the epicenter is super fun, and because everyone I know and work with is using AI all the time, it’s hard to remember that this is very much the exception.

Andrej Karpathy’s recipe for becoming an expert (opens in a new tab) has three steps: take on concrete projects and learn on demand, teach or summarize everything you learn in your own words, and only compare yourself to younger you. The middle step is exactly how I feel about working with AI.

Organizing my thoughts well enough to share them with people outside the chamber is a big part of the joy of learning here.

Natalia Quintero and Mike Taylor’s 33 Questions Executives Ask About AI (opens in a new tab) has a line I keep repeating: it’s easier to raise the ceiling than the floor.

One AI-fluent person building ten good skills and doing ten times the work is more achievable than getting ten people to each build one skill. Then you highlight the champions, and skeptics follow their peers out of practicality rather than persuasion.

That’s invitational excellence. Don’t drag the floor up. Show what the ceiling looks like and make it easy to climb.

It’s non-trivial to build a recursively learning budget and forecasting system. Every change to a category, every correction, every edit to future spending should feed back into the learning engine so next month’s forecast is better than this month’s.

It sounds awesome. It is shockingly complicated to architect. Provenance for who changed what and why, guardrails that fail closed at the database, receipts for every migration, and a review surface that can handle a few hundred line items without a modal for each one.

I’m through the hard part, I think. Ask me again in October.

A forecast is a projection of past spending into the future. A budget is a deliberate, committed behavior change that decreases expected spending to increase savings.

For years, I built budgets with a vague sense of historical actuals informing a hand wavy sense of the future. Every “budget” was really a guess dressed up as a plan, and we were often off for expected reasons. Now the forecast is transaction level certainty that tells us what happens if nothing changes, and the budget records exactly which lines we chose to change and why.

Actuals get compared to both. Drift is life not going as expected. Discipline is us not holding the line. Different problems, different fixes.

Part of the reason my email quest is floundering is because Slack is too high friction an interface surface for my email triage. To get back on track, I need a surface tuned to my preferences.

Both at work and at home, I’m thinking a lot about pulling my cron, scheduled, and loop jobs into service accounts with an agent framework.

I tried OpenClaw for two days and it was horrendous to work with. I’m trying Vercel’s Eve framework next and will report back.

Team mode with AI continues to be hard. Simply pointing people at my GTM Brain repo at work has yielded uneven results, so I’m more convinced than ever about an agent org chart in Slack that functionally operates like a highly responsive colleague.

Interestingly, the shapes we’ve drawn around work and titles for humans were optimizing for things - experience, skills, salary expectations - that have only vague relevance to agents. Learning the right boundaries for skills vs. workflows vs. spawned agents vs. named agents is a lot of trial and error.

One of the biggest unlocks for visual professionalism was our designer’s brand assets repo. Now, every prospect- or customer-facing anything uses that repo for inspo, and our polish went up dramatically overnight.

Strong recommend for every team.

I keep finding the same thing: excellent outcomes from agent work require four basic characteristics.

1. Curiosity. When I’m in doubt or uncertain, I ask the agent. What happened? What’s off? What does it see that I don’t?
2. Patience. Things will fail and break. When they do, I return to #1 until they don’t.
3. Attention to detail. I click through the work closely enough to notice what feels off.
4. Critical thought. I reason through why it’s off and how to resolve it - with #1, of course.

Attention to detail and critical thought are the critical path. One catches the issue; the other makes sense of it. Curiosity helps me resolve it.

There are skills and best practices to working with AI. I still find that every one I’ve learned is mediated through these four characteristics.

If I hit a ceiling or get super frustrated with agentic work, I check which of the four is lagging and go self-resource it.

Working with agents creates a granular fragmentation of work. Each agent carries a small stream forward, but I still carry the map of all of them.

Cycling agent tabs in Claude Code or Codex requires rapid context switching: remembering what each agent is doing, what changed, which decision is waiting on me, and where to return next. The work includes directing agents and mentally tracking multiple concurrent streams without dropping one.

That’s a very particular skill set. It’s trainable, but it isn’t intuitive.

If my agents fail, first I need to check if I’ve set them up for success.

And when they do fail, I add new guardrails + evals to make sure it doesn’t happen again and I catch it if it does.