AutoMerge is generally available

Back to Case Studies

How ENAPI Cut Time to Merge by 77% With AutoMerge

ENAPI
The Problem

With coding agents driving more than 90% of ENAPI's development work, pull requests multiplied while nearly every change still required the same human review.

The Solution

Dash0 AutoMerge scores each pull request against ENAPI's own rules, starting in dry run and moving to auto-approve after the team tuned the criteria to its codebase.

The Result

ENAPI reported a 77% reduction in median time from pull request to merge, while AutoMerge approved more than half of the team's pull requests by the second week.

The Infrastructure Behind EV Charging

ENAPI is the infrastructure layer behind the EV charging industry. It connects the companies that operate chargers with the companies that let drivers charge, and handles everything in between: data exchange, commercial contracts and financial operations.

“The shorthand I always use is that we’re the plumbing that makes EV charging work,” says Dennis Laumen, co-founder and CTO at ENAPI.

The company has 15 people, headquartered in Berlin, with a fully remote engineering team spread across Europe. Almost everyone writes code, including the operations and product teams.

Agents write the code. People still had to review it.

At ENAPI, agents drive most of the development. Engineers start almost every task in a new thread in Amp, the harness the team chose for its multiplayer mode.

“In >90% of cases, it’s all agent-driven,” Dennis says.

Before AutoMerge, AI code review agents checked each pull request first, but nearly every change still required human approval. Anyone on the team could approve a PR, regardless of its size or risk.

As agents took on more work, that process started to break. Agents unlocked the tests, refactors and side projects that had sat on engineers’ lists for months, and pull requests started coming in big stacks.

“What I’ve started noticing is that the focus becomes a scarce resource,” Dennis says.

One review process for every change also stopped making sense. “If I’m changing the AGENTS.md or adding a unit test, or fundamentally changing a flow in our system that impacts the money our partners transact, those are very different risk profiles,” Dennis says.

With agents doing most of the development nowadays, the PR review process became the bottleneck. Even though development speed increased significantly, delivery did not, because the bottleneck was in a different part of the process. AutoMerge removes that bottleneck and allows us to ship faster.

Why ENAPI didn’t build it themselves

Dennis felt the bottleneck first. The team discussed dropping reviews altogether, or hooking a model into its GitHub Actions to approve or reject pull requests.

“When you first look at it, AutoMerge sounds like quite a simple thing,” he says. “But once you start really digging into it, you see how complex it actually is.”

His advice to other CTOs who think they can build it themselves: “Don’t.”

The prompt is the easy part, Dennis says. The real work is fine-tuning it and integrating it with your own CI/CD pipeline and the risks in your repository, and that is exactly where ENAPI didn’t want to spend its engineering time.

From dry run to auto-approve in under a week

“Setup was very easy, and we were up and running within a couple of minutes,” Dennis says.

ENAPI started in dry run, where AutoMerge only posts its decision as a comment on each pull request.

“We enabled dry run mode on day one, and it built up a lot of trust in the results of AutoMerge,” he says.

The defaults were conservative. “My main complaint was: why doesn’t it approve more?” Dennis says. AutoMerge also waited for ENAPI’s AI review agents to finish and used their feedback in its decision, without any extra configuration.

After four or five days, the team switched to auto-approve. Once AutoMerge has approved a pull request, the engineer who owns it adds it to the merge queue, or asks for a human review when it matters.

“If they merge it and they made a mistake, we’re not going to blame an anonymous algorithm,” Dennis says. “It’s all the human’s responsibility.”

Using agents to tune the agent

The strongest part of ENAPI’s rollout was how it tuned AutoMerge to its own codebase. Every AutoMerge comment includes a score and a rationale for each criterion, and that text can go straight to an agent.

Dennis had his agents analyze past pull requests and AutoMerge’s comments, refine one criterion at a time, save a new configuration in Dash0, and replay it on past pull requests in the Testing tab.

It’s basically just a loop. You feed it in there, do a test run, get the results from the test run, feed it back into your agent, refine again, and go from there.

AutoMerge also recognized risky code on its own. “It kind of knows which code is risky as well,” Dennis says. When a change handled money, AutoMerge held it for a person. ENAPI reinforced that in its criteria, for money flows and core EV charging functionality.

The same loop improved how ENAPI’s coding agents write pull requests. Dennis added guidance on what AutoMerge looks for to the team’s agent wiki and tested whether pull requests got approved more easily.

“It’s a nice reinforcement loop where it just pushes you towards best practices,” he says. “Keep your PR small, make sure the blast impact is limited, make sure you provide evidence.”

More focus where it counts

Changes like agent guidance updates, additional tests and small refactors often got routine human rubber-stamp approvals. Now AutoMerge can approve the ones that qualify, so engineers can focus on the changes that need closer attention.

AutoMerge has allowed our engineers to focus on shipping end-user value to our customers instead of spending all of their time on reviews.

For Dennis, that also lowers risk. When engineers aren’t buried in routine reviews, the critical pull requests get the attention they deserve.

His pitch to other engineering leaders:

Let humans review the changes that actually matter and let the agents handle the rest.
  • Share