What running slow taught me about agentic engineering

Hi friend!
​
AI tools can tempt you to work at 100x speed.
​
And honestly, I’ve been there.
​
Spinning up ten agents in parallel. Jumping from tab to Codex to terminal to Cursor like a madman. Moving fast. Getting a lot done. Seemingly.
​
Then reality kicks in.
​
Your mind can only go so fast.
​
And if your goal is not just to generate output, but to actually ship something good, you cannot simply hit send on everything the AI gives you.
​
You have to review it. Understand it. Challenge it. Spot the holes. Sometimes take two steps back.
​
Very quickly, all that speed can leave you with surprisingly little to show for it.
​
I think there is a better way.
​
A few days ago, I was out for a run. For the first time, a voice inside my head said:
​
Go slow.
​
So I slowed down.
​
I glanced at my Apple Watch and saw my pace drop below a “magic” 6 min/km mark. At first, it felt wrong. Like I was underperforming.
​
But then something shifted.
​
I reached a pace where I felt like I could keep going forever.
​
I had read many times that slow running builds the capacity to go far. But my ego always got in the way. I kept chasing some arbitrary “fast” pace, even when it was not helping me become a better runner.
​
And because running always seems to unlock better thinking for me, the analogy suddenly clicked.
​
This is exactly how I should work with AI.
​
Why try to go as fast as possible when the goal is to go far?
​
So I started slowing down in my agentic coding work.
​
One task at a time. One thread of thought at a time. Actually reading what the agent produced. Challenging its assumptions. Staying close enough to the codebase that I could still understand what was changing.
​
And suddenly, I felt like I was achieving more.
​
The code was better. My reviews were faster. I had fewer messy detours. I did not lose the plot halfway through the work.
​
Going slower helped me go further.
​
Maybe that is the more useful version of the “100x developer” idea.
​
Not 100x speed.
​
Maybe it is 10x speed with 10x quality.
​
That still gets you to 100x.
​
Thanks for reading,
Robert

Headlines

• ​Claude launches Dynamic Workflows, a way to let Claude coordinate large, long-running tasks with multiple parallel subagents that delegate, and verify subtasks dynamically. Similar to Cursors /orchestrate feature
  ​
• ​Codex Computer Use now also works on Windows
  ​
• Clicky is working on making computers truly controllable with voice:

Farza 🇵🇰🇺🇸 @FarzaTV Watch me control my computer with just my voice. This is the future of operating systems. No hands. GPT-Realtime 2.0 is very, very underrated. Demo: 11:26 PM • May 30, 2026 792 Retweets 13695 Likes Read 896 replies

​

Miscellaneous

• ​A conversation with Cursor's Federico Cassano on how Composer 2.5 was trained​
  ​
• ​I made a little Harness-explainer post on LinkedIn​
  ​
• Cursor's Lee Robinson shares helpful tips on making it easier for your agents to understand your codebase:

Lee Robinson @leerob Some tips to help agents understand your codebase: 1. The source code either needs to be the source of truth, or have something legible as a path to the source. For example, if marketing site content is actually stored in a CMS, you need to either delete the CMS and move that 9:17 PM • Jun 1, 2026 43 Retweets 750 Likes Read 20 replies

​