Builder First. Operator Always. | Former Founder | Write about AI, product leadership, and other fun topics on @Substack | Love being active: 🏂 🚴 🏈 ⚽️ 🏌️♂️
- Tweets 2,006
- Following 312
- Followers 150
- Likes 1,582
Pinned Tweet
I want to connect with more founders and AI-native operators. Who’s on here? Drop a comment and say hello. 👋 😎🔥
8
1
10
728
If a team is spending $10K a month on AI and only looking at seat count or usage logs, they are measuring the wrong thing. The number that matters is whether the team ships more, cuts handoffs, and can point to output that covers the bill.
1
1
9
Agentic security gets useful when it can close the loop all the way to a patch. Finding bugs is the easy part. Reproducing them, generating a fix, and getting a maintainer to trust the result is where the work starts.
7
A lot of company work is moving into agent loops before it ever reaches a meeting.
Who owns the workflow, who reviews the output, who overrides the agent. That starts to matter more than who typed the first draft.
1
2
11
The brag changed.
Nobody cares that you can call a frontier model anymore. Everyone wants to know if you can make it cheap enough to ship at margin.
Routing, batching, caching, retries, fallbacks. All the ugly ops work that turns a demo into a business.
1
1
23
The browser devtools tab is turning into a real runtime for agent workflows.
Agents inside the place developers already inspect, debug, and operate from. Faster to adopt. Less ceremony.
3
3
99
Self-repairing software starts to matter when it can make the fix, verify it, and ship it without a human PR loop.
Fewer broken weeks. Less support drag. Less engineer time burned on repeat maintenance.
3
13
Creator-led growth works when the product gives people something worth showing off.
If the best ad is a clip of someone using your product, the creator is doing distribution and product proof at the same time.
Pay them well enough to keep shipping. That's the moat.
1
2
21
If models can reproduce bugs and push fixes directly, PRs stop being the default unit of contribution.
The better product is the one that lets the person hitting the problem repair it in place.
1
2
20
I like agent setups that stop at a human review point.
Once the thing is chaining shell, edits, tests, and formatting in one blur, you lose the plot fast. Supervision matters more when the workflow can break in five different places.
1
3
15
Granular permissions are the product.
The agent layer gets useful when each connector can ask for one narrow action, get human confirmation, and stay away from reset emails and inboxes.
If it can read everything, it will do the wrong thing faster.
1
8
Jeff retweeted
Very fond of the people who turn a quick hello into a conversation worth coming back for 💛
Thank you: @AllenTheDetails @annalea_l @justinhammon_ @klausbyskov @klein_more88816 @Zenul_Abidin
8
3
20
589
Standalone agent apps win when they remove the typing tax across the whole computer.
Once you use one long enough, opening menus, copying text, and bouncing between apps starts to feel like manual labor.
That is the product. Not the demo.
9
The best product in a category is often the one that changes the category's job.
Another version can have cleaner craft, sharper UX, better copy, better direction.
The benchmark is the thing that resets expectations.
That is why legacy and quality are different measurements.
6
AI infra is getting weird in the same way telecom got weird. Too many intermediaries. Too little direct ownership of the buyer.
The winners will own the workflow people run and the relationship that tells them what to provision next.
8
Cheap inference only matters when it changes what ships. If the model is already quantized at the provider layer, the real edge is distribution, workflow fit, and getting output that is good enough to turn into a product people keep using.
5
The cheap model race is really a workflow race.
If memory setup, routing, evals, and repeat deploys are painful, teams stop experimenting. The stack that wins makes those pieces boring enough that nobody argues about them anymore.
1
1
12
AI code is moving work downstream.
Less typing. More review. More paging. More incident response when a fast merge shipped something nobody had time to really read.
If your team only measures PR throughput, you will call it progress for a while.
The bill shows up in ops.
1
4
Safety monitoring is becoming part of the model stack.
If you can catch weird behavior before it becomes a customer incident, you can ship agents people will trust and pay for.
1
7
Screenshot-to-action is the first assistant workflow that feels native.
No hunting through menus. No recreating the context by hand. You send the image, say what you want, and the assistant handles the thing from there.
That is where the real product shift is.
4