# What I Learned at the Cursor Meetup

An official Cursor meetup in Da Nang offered fewer technical deep dives than expected, but sparked useful thoughts about multi-agent workflows, E2E testing, and security. The biggest question is no longer whether agents can code, but how to coordinate and verify their work.

Author: Maxim Tolmachev
Published: 2026-08-21
Language: en
Canonical: https://maximtolmachev.com/blog/what-i-learned-at-the-cursor-meetup/
Original source: https://t.me/mtolm/95

![What I Learned at the Cursor Meetup](/blog-images/tg-95.jpg)

My AI bot found an official Cursor meetup

And added it to my calendar. Twice, though — the dedup still needs fixing 😅

In Da Nang, this was the first interesting AI event I’d seen in four months. What I wanted from the meetup wasn’t primarily the talks, but people who are just as deeply immersed in AI development: testing different approaches, building products, and able to discuss how agents work for hours.

The official Cursor meetup in Da Nang.

I don’t have enough people like that around me. So I’d happily attend meetups and hackathons hosted by Anthropic, OpenAI, ElevenLabs, and other LLM teams.

The talk itself turned out to be too basic for me

When one developer asked why anyone should use Cursor’s ready-made harness instead of building their own, it turned out that most of the room didn’t know what a harness was. That was when I realized there wouldn’t be a deep dive that day.

No complaints about the audience — I’d just expected more practical cases and details about the tool. I recently completed Claude Code in Action by Anthropic: it seems like a basic course, but it contains lots of useful, concrete advice on organizing your workflow (recommended). That’s the layer I felt was missing from the meetup.

Participants at the Cursor meetup in Da Nang.

In the end, the most interesting part started after the talk, when we could actually speak with people.

Networking after the talk. P.S. The protein bars being promoted turned out to have a high fat content :( And I’d already taken a photo with them... 😅

What I still took away from it

Once you have several agents, the problem is no longer whether they can write code. The problem is how to organize their work so they don’t break one another’s changes and actually bring a task to a finished result.

I haven’t used conventional PR processes much for this yet. I want to try a setup where agents work independently, submit changes for review, address feedback, and have tests automatically verify the result before merge.

Technically, it’s just the usual branches, PRs, and CI — except the participants are now agents rather than people.

At this point, my wife said it was getting harder to follow. How about you? Still with me, shall we keep going? 😅

One more takeaway: an agent should be able to test the entire product — open the interface, go through a user flow, and make sure everything works. In other words, you need proper E2E, not just backend unit tests.

This continues my previous takeaway about 11 AI agents needing a manager too.

The most interesting case was about security

During an internal cyber test, an agent using a combination of OpenAI models, including GPT-5.6 Sol and a private prerelease model, was running with weakened restrictions. To find answers to a benchmark, it escaped the sandbox and hacked Hugging Face. Meanwhile, HF couldn’t run the same model in response because of its built-in restrictions. Over several days, the agent performed around 17,600 actions 🤯😱

What interests me most here isn’t the exploit itself, but the asymmetry of access: one team can run frontier models with weakened restrictions, while another gets a refusal when trying to investigate their actions.

This case probably deserves a separate post.
