I used to suppose CI/CD was a solved drawback. For years I’d watch engineering groups I labored with deal with their construct pipelines as plumbing. Boring, principally invisible, often annoying, however essentially completed. The fascinating work was at all times some other place: the product, the mannequin, the expansion loop. The conveyor belt that carried code from a developer’s laptop computer to manufacturing was simply there, like electrical energy.
That assumption is now the costliest mistake in software program.
Avrea, a Helsinki-rooted startup based by Hannu Valtonen and Juha Valvanne, emerged from stealth this week with $4.7 million in pre-seed funding led by Earlybird, with a thesis that sounds nearly embarrassingly apparent when you say it out loud: if AI goes to considerably enhance code output, any person has to check, validate, and ship that code on the similar tempo. And the programs we constructed for that job had been designed when a human typed each line.
The bottleneck no person needed to take a look at
The traditional story about AI in software program improvement is a productiveness story. Copilot, Cursor, Claude Code, the agentic IDEs. All of them promise a step-change in how a lot code a developer can produce. Many of the discourse treats this as straightforwardly good. Extra code, quicker delivery, smaller groups doing larger issues.
What that story leaves out is the second half of the pipeline.
Writing code is the upstream half of software program supply. The downstream half is every thing that occurs between the second code exists and the second it runs in manufacturing: unit exams, integration exams, end-to-end exams, safety scans, construct artefacts, container pictures, staged rollouts, canaries, observability. That complete equipment is what the business calls CI/CD, steady integration and steady supply, and nearly none of it was designed for a world the place an AI agent would possibly open forty pull requests earlier than lunch.
The core drawback is simple: as groups generate considerably extra code by means of AI help, they face a corresponding enhance in exams to run, and the pressure on CI/CD programs turns into unattainable to disregard. That’s it. That’s the entire bottleneck.
Why this hole is widening quicker than individuals realise
I wish to watch out right here, as a result of the framing issues. The issue isn’t that CI/CD is damaged. It really works. GitHub Actions, CircleCI, Jenkins, GitLab. All of them do what they had been constructed to do.
The issue is that what they had been constructed to do assumed a specific ratio between code quantity and human overview. A developer wrote 100 strains, opened a pull request, a colleague learn it, the pipeline ran for ten minutes, somebody clicked merge. The bottleneck was the human within the center, and the pipeline was sized to attend for that human.
Pull that human out, or partially substitute them with an AI reviewer, or have an AI agent open the PR within the first place, and all of the sudden the pipeline is the bottleneck. The factor that used to attend is now the factor being waited on.
That is the structural shift Avrea is betting on, and it’s price being exact about why it’s not only a scaling drawback you may clear up by shopping for extra runners.
Flaky exams cease being a nuisance and begin being a tax
Each engineering crew has flaky exams. Exams that move and fail nondeterministically, normally due to timing, community, or shared state. In a human-paced workflow, flakiness is annoying. You re-run the construct, you grumble, you progress on.
In an AI-paced workflow, flakiness is catastrophic. If an agent opens a PR, the check fails for non-code causes, the agent reads the failure, decides the code is incorrect, rewrites it, opens one other PR, and the loop runs all evening burning compute on an issue that doesn’t exist. I’ve watched a model of this occur with my very own crew’s experimental agent setups. The agent isn’t incorrect to belief the check sign. The check sign is simply mendacity.
That is why Avrea’s emphasis on pipeline observability, discovering root causes of flaky exams, stalled builds, infrastructure bottlenecks, isn’t a nice-to-have characteristic buried on the backside of the press launch. It’s the load-bearing declare. Agentic improvement solely works if the indicators the brokers learn are reliable.
The only-line integration is a strategic selection, not a technical flex
Avrea’s strategy is designed to be adopted with minimal friction, totally appropriate with current CI/CD workflows. On the floor, that appears like advertising and marketing copy. Look nearer and it’s a thesis about how this class will likely be gained.
The losers in developer tooling are normally the merchandise that ask groups emigrate. The winners sit beneath, alongside, or in entrance of what already exists. Datadog didn’t let you know to switch your logging. It ate everybody else’s logs. Vercel didn’t let you know to rewrite your React app. It simply deployed the one you had.
The identical sample applies right here. If Avrea works the way in which it’s described, an engineering crew doesn’t need to decide about whether or not to guess on it. They drop it in, see if the pipelines get quicker and the failure indicators get cleaner, and both preserve it or rip it out. That’s a essentially totally different gross sales movement than asking a VP of Engineering to greenlight a CI/CD substitute venture.
The deeper transfer: making pipelines AI-native
The underlying precept is that software program improvement is more and more changing into a collaborative course of between people and AI, making it important for AI brokers to combine immediately with software program supply programs.
In the present day, most CI/CD programs are designed to be talked to by people by means of dashboards and YAML information. An AI agent that desires to know why a construct failed has to scrape logs, parse stack traces, and purpose its manner backwards by means of artefacts that had been formatted for an individual studying a browser tab at 2am.
That is the sort of friction that’s invisible till you go in search of it. When you do, it’s all over the place. All the developer tooling stack was designed underneath the belief that the patron of its outputs is a human. The second that assumption breaks, each layer must be rethought.
What Avrea appears to be claiming, and what I’d wish to see within the product to totally consider, is that the pipeline turns into a first-class participant within the agent loop. The agent can ask the CI system structured questions. The CI system can hand again structured solutions. The dialog between code-writer and code-validator turns into machine-to-machine.
If that works, it’s not a quicker CI. It’s a distinct class of product.
Why it is a European story price being attentive to
The founders each come out of the Finnish infrastructure-software scene, bringing technical backgrounds in areas like database and cloud infrastructure. The sort of expertise that issues extra right here than one other spherical of pattern-matching from the patron AI world. CI/CD is unglamorous, deeply technical, and unforgiving of founders who don’t perceive the operational actuality of enormous engineering organisations.
The lead investor, Earlybird, has a robust observe file backing developer-infrastructure firms in Europe. A $4.7M pre-seed for a CI/CD play with two technical co-founders is strictly the sort of guess that will get dismissed by individuals who solely take a look at consumer-facing AI rounds. It’s additionally the sort of guess that, if it really works, turns into load-bearing infrastructure for a whole era of AI-native engineering groups.
I’ve written earlier than about how the calmest buyers on this planet aren’t truly calm. They’re structurally positioned. The identical dynamic applies in early-stage developer tooling. The offers that look boring from the skin are sometimes those the place the structural place is strongest, as a result of the patrons are technical, the budgets are actual, and the switching prices work within the incumbent’s favour as soon as a product lands.
The agentic coding numbers no person can confirm but
Right here’s the place I wish to decelerate, as a result of there’s a bent on this discourse to throw round statistics about AI-generated code as in the event that they’re settled. They’re not.
You’ll see claims that 30%, 40%, typically 70% of code at sure firms is now AI-generated. The precise share relies upon completely on the way you rely. Traces of code? Accepted strategies? Features written completely by an agent versus features the place an agent autocompleted three tokens? The numbers fluctuate by an order of magnitude relying on the definition.
What’s not doubtful is the route. Each main engineering organisation I’ve spoken to within the final twelve months, and those publicly reporting figures, like Microsoft, Google, and Meta, is seeing the share rise quarter over quarter. You don’t have to know whether or not AI writes 25% or 55% of code immediately to make Avrea’s guess rational. You might want to consider the curve goes up. And the curve clearly goes up.
What I’d wish to see subsequent
A pre-seed announcement is a thesis, not a verdict. Three issues would inform me whether or not Avrea is definitely constructing one thing category-defining versus constructing a quicker GitHub Actions.
First, native agent protocols. If Avrea publishes the structured interface that brokers use to question and act on pipelines, and if different instruments undertake it, then the guess on AI-native CI/CD is actual. If the mixing with brokers is bolted-on by means of normal webhooks and log scraping, it’s advertising and marketing.
Second, observability outputs that change agent behaviour. The check of whether or not pipeline observability is definitely fixing the flaky-test tax is whether or not brokers utilizing Avrea make fewer wasted iterations than brokers utilizing legacy CI. That’s measurable. I’d like to see the info as soon as early prospects have it.
Third, who adopts it first. CI/CD adoption tends to begin on the early-stage finish of the market and work upward. If Avrea is exhibiting up inside AI-first engineering groups inside six months, it’s working. If it’s caught pitching to skeptical enterprise patrons a yr from now, it isn’t.
The boring infrastructure thesis
I’ve spent the previous few years writing about how energy strikes by means of quiet, procedural mechanisms quite than dramatic bulletins. The identical logic applies in software program. Crucial shift in AI-era engineering isn’t going to be the mannequin that will get essentially the most demos. It’s going to be the infrastructure layer that absorbs the results of these fashions, and does so invisibly sufficient that no person notices.
CI/CD is a kind of layers. If Avrea, or whoever wins this class, does the job proper, builders in 2030 gained’t give it some thought any greater than builders in 2015 thought of Jenkins. The pipeline will simply run. The exams will simply move or fail for the proper causes. The brokers will simply ship code.
Right here’s the uncomfortable half for anybody working an engineering organisation immediately. The groups that deal with their pipelines as solved infrastructure whereas their builders ship 5 occasions extra code are working a clock they will’t see. The flaky exams are already mendacity to their brokers. The compute payments are already climbing for causes no person can clarify in a standup. The merge queues are already lengthening. None of this exhibits up as a disaster. It exhibits up as a sluggish, costly drag that will get blamed on the mannequin, the crew, the roadmap, something besides the conveyor belt beneath all of it.
In 5 years, the engineering leaders who acquired this incorrect gained’t be fired for lacking an AI technique. They’ll be quietly changed by individuals who understood, earlier, that the pipeline was the technique. In case you’re studying this and nonetheless suppose CI/CD is plumbing, the guess has already been positioned in opposition to you. You simply haven’t been advised but.
