agents and ephemeral deployments
or, how to get sh*t done while driving your 1.5hr commute
I spend a lot of time driving - on average 9 hours a week. In between podcasts and discordant car acapellas (a-car-pellas) is a lot of thinking, reflecting and ideating. In the span of just 13 months, my phone has entirely filled up with voice memos. Despite that, I have little to show for it because that time is generally in-actionable. I like having a reserved space for thinking but, at the end of a long drive and a longer workday, I lack the bandwidth to materialize most of them.
In the past couple months, two new developments have changed my ability to act on those ideas. The first, is that I started using Vercel. The second, is that Cursor Cloud Agent is starting to get really useful.
Over the last year, I’ve retooled most of my projects to support a Vercel+Lambda stack. This is motivated by the fact that a) I’m building more small exploratory projects than ever before (yay!), and b) my AWS bills were, at one point, $60/month (hint: it was the ALB). As of January ‘26, they’re back down to $5. Phew.
Converting to Vercel is difficult; building on Vercel is easy. Making a generalizable template with a preconfigured stack is bliss. I can deploy up a new project in a couple minutes and now, with cloud agents, I can build it all in a half-hour.
The key unlock is ephemeral deploys for PRs. My typical workflow requires pulling the particular branch and testing locally. Now, I can manage all of it on my phone - between dictating instructions to the agent, to inspecting results on the ephemeral deploy, and reviewing code in the PR.
An optimal two-way commute could look like
get an idea (usually from a podcast) -> market research using XYZ deep search -> iterating on validation/wedge
<9-7 workday>
fork template -> dictate to cursor -> inspect results when I get home
Thinks to keep in mind:
- You must build robust cursor rules - it is not a good idea to review code minutiae while driving.
- If you've got Clawdbot/Openclaw setup, I've found some success asking it to summarize code-structure at a high-level and use the Whisper plugin to explain it to me. Some success.
- Style guides/communicating product feel and taste. This is just a challenge with using AI for small projects that don't have component libraries IMO.
- Testing.
- My colleague Swiz recently wrote a cool dive into AI-driven UI testing. A quick test (note: I used OpenRouter since I don't have OAI premium) was cost-prohibitive for personal projects, though. Agents aren't token-efficient when navigating browsers yet.