10x'ing our eng velocity with The Zerg

As a (very) early stage startup, speed and efficiency are our biggest advantage. Though we’re always trying to execute on all cylinders, with the release of Claude 4, we’ve made a huge breakthrough.

Ship Happens (Even When We’re Not Looking)

We now have a swarm of agents that are 10x’ing our build speed. We can build while we’re commuting, on a date, sleeping, building other stuff, you name it. As quick as models have been improving, I’m constantly surprised at how much more leverage I have now.

Here’s a recent example. Last week, we had a dramatic increase in the amount of data we were processing and the site was basically unusable. To resolve it, we had to overhaul our data fetching strategy across the stack for multiple paths. This was a pretty complex task requiring new materialized views, optimized indexes, batched calls, rebuilt views with paginated requests & infinite scrolling, offloaded processing to asynchronous threads, etc. Normally, this would take a single dev at least a week.

With our new tooling, I shipped a complete rewrite of our data handling across the UI, backend, and database in a single day. Overnight, we’re meeting our new scaling requirements on an optimized stack.

Enter The Zerg

At a high-level, The Zerg are a swarm of Claude 4 instances that live across Linear, Github Actions, and our terminals. Along with MCP servers and additional infra we’ve built to let them talk to each other, they deeply understand our code, independently plan tasks, break them down into tickets, and commit high-quality PRs.

We’ve been using Cursor, Windsurf, Codebuff, Claude Code, and several other codegen tools since their early days, but with a few key improvements and the release of Claude 4 we’re now building at a speed that puts our former pace to shame. Things aren’t yet perfect (probably <20% of PRs are zero-shot successes) but with some guidance they’re remarkably good.

The last few weeks has really felt like living in the future, and we feel confident that all product development teams are going to adopt some shape of the same process, optimized for their codebases, infrastructure, and organization. As models inevitably continue to improve, orgs are running out of excuses - they’ll need to adapt or die.

The Future of Work

So what does our process look like now? There are 3 main patterns we use, in increasing level of complexity and supervision:

1) For smaller one-off tasks like product quality issues, UX fixes, changing graphics, etc; now we can simply create a Linear ticket with a basic, one-liner description, assign it to the Zerg, and within 5-10 minutes have a PR ready for review.

2) For more complicated tasks, we’ll plan out the task in a back & forth conversation with Claude Code; since it has the context of our entire codebase, we can quickly talk through the desired outcome, implementation options and strategies, acceptance criteria, and even what not to do. Then it breaks down the task into independent tickets, which it automatically creates in Linear with high-quality documentation, and assigns them to…The Zerg. An individual instance will pick up each ticket, and implement it with proper knowledge of dependencies, build order, and more. The final set of PRs are then ready for review.

3) For more core, complex, or wide-ranging changes, we actively pair program with Claude Code, walking it through what to implement and how, and testing as we go. This is most similar to our previous workflow with Cursor or other coding agents. We'll employ pattern 2 above when possible to accelerate through it.

Once PRs are ready for review, we usually pull them into a local branch to test things and ensure it works as expected before merging. This is intentionally manual for now, but we have plans to automate even more of this soon.

What’s Next?

We’re planning several improvements speed things up even further, from automated testing to better integration with our product context - our goal is to ensure that future hires are immediately able to build 10x faster as a result of our system.

If you want a full stack write-up, drop us a line at founders@edgedive.com!