GitHub is the wrong shape for this new world

Jul 30, 2026 05:27 AM - 8 hours ago 3

A batch of attraction is fixed to the wide capacity and reliability of GitHub. Rightfully so. But I deliberation the much absorbing point to mobility is the paradigm that we usage pinch GitHub. It's the incorrect style for really we build package today, and we request amended tools, workflows, and infrastructure primitives that meet the caller demand.

Software engineering has changed

I recognize this connection is simply a batch for illustration saying "water is wet."

We unrecorded successful a fundamentally different world for processing package today. Are location acquainted things? Definitely. Are location things from the past that we tin usage successful this caller world? Absolutely. Code is still codification aft all. As a mentor utilized to preach to me, it's bytes successful and bytes out.

The bottlenecks we underinvested successful are still around, from root power to CI/CD to codification reappraisal and moreover deployments. Only now they are sidesplitting the newfound velocity we have.

The symptom of these bottlenecks is compounding because everyone tin now lend code. Code has expanded extracurricular of an engineering squad and into different teams for illustration sales, support, and marketing. Meaning, a 10-minute build is felt by virtually everyone successful a company, not conscionable a mini subset.

This has profound implications, arsenic it calls into mobility the assumptions and things we person accepted arsenic normal. It raises questions that I deliberation are still unanswered:

  • Who owns the code? Or, put different way, who is responsible for the codification that gets deployed?
  • How do we spot the code?
  • How do we reappraisal each of this caller codification being generated?
  • How do we link codification from 1 spot to activity pinch codification successful different place?
  • How do we measurement the costs of token walk pinch the output of value features?

Ask these 5 questions of 5 different people, and you'll astir apt get 5 different answers.

This is what I mean erstwhile I opportunity that package engineering has changed. It's not conscionable that LLMs and agents tin make an full characteristic disconnected a five-sentence prompt. That's a instrumentality we are using.

It's that the full paradigm astir package engineering is now drastically different and moving astatine a velocity that we simply can't support up with. Assumptions we person made are crumbling pinch each passing release.

I judge we must rethink our paradigm and the devices we are using. We are bending our existing devices and human-centric paradigms into this caller world. This is the incorrect approach.

Collaboration vs infrastructure

Really, this blog station should beryllium titled thing for illustration "GitHub, GitLab, Bitbucket, and others are the incorrect style for this caller world." But that's a mouthful, truthful erstwhile I opportunity GitHub, publication it arsenic each of them.

I deliberation of GitHub arsenic a collaboration tool. Google Docs is simply a spot wherever I tin spell to collaborate connected this blog station pinch our method penning team. GitHub is simply a spot I tin spell to collaborate pinch different engineers connected the codification that makes up Depot.

I grew up pinch GitHub. In fact, I wanted to activity astatine GitHub erstwhile I first became a package engineer. My full musculus representation is built astir the paradigm that GitHub pushed maine to use.

  • Write codification successful a branch.
  • Open a propulsion petition erstwhile I'm fresh for my changes to beryllium reviewed.
  • Wait for CI and checks to walk to validate that everything works.
  • Review comments from my colleagues and spell backmost and distant connected ideas successful the PR.
  • Once everything is greenish and my colleagues person signed off, merge my changes.

There are derivatives of this workflow. There are wildly differing opinions connected the effectiveness of this paradigm. But inquire immoderate engineer, and they will cognize it for illustration the backmost of their hand.

When LLMs first started to travel online, it was earthy to bolt them onto our existing paradigm. They could constitute codification successful a branch, unfastened a propulsion request, get a codification reappraisal from a quality (or different agent), usage the PR to collaborate via comments, tally our tests via our existing CI pipelines, and yet merge the code.

It was logical. The systems existed, and the agents astatine that clip still mostly moved astatine quality speed.

Then everything changed. Agents pinch the latest procreation of models sewage incredibly good. Freakishly good. We went from models wherever we often had to nudge the Jenga blocks backmost into spot to get bully codification out, to models that nutrient consistently correct codification fixed capable context.

What's the nett consequence of amended models? More code, much branches, much parallel work, and much strain connected our existing clunky human-powered paradigm.

Our collaboration pattern, and the underlying systems that backmost it, are now the superior bottleneck. Every engineering squad is emotion bottlenecks successful GitHub itself, CI, codification review, propulsion requests, information scanning, and moreover deployments.

There is now an impedance mismatch betwixt our devices and really we want to build. Our existing collaboration paradigm, mostly built astir GitHub, is now successful conflict pinch really we really build package today.

We must rethink our paradigm for package delivery. Not done the lens of human-centered collaboration, but done high-throughput infrastructure primitives.

High throughput package delivery

Building a caller package transportation paradigm requires taking a measurement backmost from the human-centric process that we person utilized complete the past decade. Rethinking package delivery, not arsenic a series of quality actions, but arsenic a continuous automated process driven by systems.

Once you commencement imagining that world, the problem of package transportation starts to look little for illustration a collaboration instrumentality and much for illustration infrastructure. Where infrastructure, successful this context, intends a mini number of primitives connected which everything other tin beryllium built. We tin look to really we person adopted cloud-native technologies for inspiration:

  • Compute arsenic a primitive gave america VMs
  • Storage gave america entity stores
  • Networking gave america load balancers

We should beryllium reasoning astir the basal infrastructure primitives we request for package delivery:

  • Source control: durable, immutable history of really codification evolved
  • Execution: isolated, performant compute for building, testing, and validating changes
  • Artifacts: reproducible outputs that tin beryllium moved crossed systems
  • Caching: reuse of deterministic work
  • Identity: proving who aliases what produced code
  • Policy: machine-enforceable rules for quality, security, and compliance

These aren't features of a developer tool. They're infrastructure primitives. Just arsenic we stopped reasoning astir provisioning individual servers and started reasoning astir compute, storage, and networking arsenic composable building blocks, package transportation needs the aforesaid shift.

The winners of the adjacent decade won't build a amended propulsion request. They'll build the infrastructure that makes package generation, validation, and deployment run astatine instrumentality scale.

Related posts

  • What we request from CI for agentic engineering
  • The bottleneck has shifted from penning codification to integrating it
  • Staying successful power of your codebase successful the AI era
Kyle Galbraith

Kyle Galbraith

CEO & Co-founder of Depot

Platform Engineer who despises slow builds turned founder. Expat surviving successful 🇫🇷

More