Dharmesh Shah

HubSpot

Founder of HubSpot, a venture-backed software startup.  HubSpot offers the industry's first inbound marketing system for small businesses.  The software is available as a hosted service (SaaS).

Founder and CTO

OnStartups.com

Popular blog and community for startups with over 900,000 subscribers.

Founder

Pyramid Digital Solutions

Founder and CEO of Pyramid Digital Solutions, an enterprise software startup in the financial services sector.

Pyramid was acquired by SunGard Business Systems, an $11 billion technology company, in August, 2005.

Founder and CEO

SunGard Employee Benefit Systems

Software Developer

Founder and CTO at HubSpot. Helping millions grow better.

Career startup guy.

Author of the book &quot;Inbound Marketing&quot; (http://InboundBook.com)

Specialties: Entrepreneurship, Software Development, internet marketing,SEO

I've switched almost entirely over to agentic coding (what most call "vibe coding").

A few observations:

1) It does reduce the amount of typing I have to do. My wrists are thankful.

2) I do get tasks done much faster than I would otherwise. But, while AI is doing the work, the tasks run longer, and I can't really do anything productive. So most of the time, I'm just casually watching the work that it's doing in the terminal.

Related thought:  If Claude Code, OpenAI Codex and others offered an "Ultra" tier for $2,000/month (10x the cost), but reduced latency by 10x, I would pay it. It's economically valuable to me to not have to sit and wait at each step.

3) Though the code generated is not as good as I would have written manually myself, there is an upside: I tend to write more unit tests, tend to refactor more frequently. It just makes me more likely to take the "good, clean, living" path when I don't have to do all the work. It's not that I'm lazy, it's that I'm impatient.

4) I'm using Tailwind CSS again (which I had moved away from) because I don't have to really edit the Tailwind code. This also gives me the ability to use shadcdn UI components (which use Tailwind).

5) My biggest feature request is a better way to review and edit the generated code. Cursor is a full IDE (based on VS Code) but still uses mostly a file-based editing mechanism. What would be more useful is visual "blocks" (maybe even presented as a tree of blocks) with what changes were made for this session. Then I could click on one of those blocks and edit that block of code. Or, right click on that block and run a command with that block as the context.

Example: System generates a block of code in a function somewhere. I should be able to right-click that code and prompt: "Turn this into a centralized function in a new module named XYZ".

6) As far as the models go, it feels like OpenAI and Anthropic are at parity with each other. Both great models. The models are not the issue now -- it's more about the UX and ergonomics.

7) I know Anthropic has a 1M token context window now. If I were them, I'd update the UI of Claude Code to *show* that number somewhere in the UI (probably where they show the percent of context used before next cleanup). Right now, I have no way of knowing if I'm using that 1M token widow.

8) The agents should see patterns in the kind of revisions/edits I ask for and if I ask for it multiple times, perhaps offer a suggestion to update the CLAUDE.md (or equivalent) file. Right now, I have to manually come up with what to put in there as I go. 

9) The agents should do a better job of looking at what I'm trying to do and have suggest existing libraries. It feels like it just defaults to write the code that I ask for, instead of saying: "Should I use XYZ library?"

10) Feels like there should be better native support for web app development in terms of being able to run the web app, interact with it, watch the Javascript console for errors, etc.

Today:  BYOD (Bring Your Own Device)

Tomorrow: BYOA (Bring Your Own Agents)

More and more people will be building their own agents to increase their personal productivity. And, they will take that "team" with them to their next job. 

I can also imagine a day where companies will gravitate towards hiring people that have their own team of agents that help them create and deliver more value.

How to use AI:

To test your thinking.

To clarify your thinking.

To elevate your thinking.

How NOT to use AI:

To replace your thinking.

~ From my upcoming keynote at #INBOUND25 conference in San Francisco in almost exactly two weeks.

Taplio

Dharmesh Shah's Linkedin Analytics

What is Dharmesh talking about?

Who is engaging with Dharmesh

's Best Posts (last 30 days)

Want to drive more opportunities from LinkedIn?