Atlanta in August is a commitment (and not the kind you can git revert). You pack for humidity, you plan your walk from the hotel to AmericasMart, and you accept that the weather is part of the experience.

What you get in return is RenderATL: two days, one badge, 8,000+ people, and a schedule spanning a main stage keynote, an AI expo, the first Node.js Interactive ever held in the South, a career fair, a hackathon, and a music festival that runs into the night.

We came for the Markdown, which we admit is not the most exciting sentence ever written about a tech/music festival. Stay with us. Four talks, taken in order, added up to an argument we do not think anyone planned.

The week around RenderATL 2026

RenderATL sits inside Atlanta Tech Week, and the branding is everywhere: Building Atlanta Together, pixel type against circuit-board line work, on every hallway backdrop in the building.

A lot of tech conferences are held in a city. This one is of a city. The programming, the speakers, the music, the sponsors, all of it points back at Atlanta as a place that builds software rather than a place that hosts an event about software. If you have only experienced the Bay Area version of a developer conference, this one recalibrates something.

Justin E. Samuels: the case for building it here

RenderATL’s founder and CEO opened the conference and did the thing founders on their own stage almost never do well: he made the numbers into an argument instead of a victory lap.

The numbers are worth stating plainly. RenderATL started in 2019 with 300 attendees at what was then the South’s first React-focused conference. This year it filled more than 250,000 square feet with over 8,000 attendees, 150-plus speakers, and 70-plus sponsors and partners, with Zillow as title sponsor. Across the wider Atlanta Tech Week, more than 15,000 people took part. Cumulative economic impact on the city: north of $55 million.

Justin’s framing was not “look how big we got.” It was closer to: this was always possible here, and the only reason it had not happened was that nobody had done it.

He has been consistent for years that people are more than their job titles, and that the conferences he actually remembered were never really about the keynote. You can see that philosophy in the architecture of the event. Render is deliberately not a pure engineering conference — it is a technical conference with a music festival, a career fair, and a hackathon bolted to it, because the assumption is that the person in the room is a whole human with a career, a community, and taste.

For a room of engineers, most of whom will never run a conference, the transferable part was the argument about place. You do not have to leave to build something that matters. You have to decide the thing is worth building where you already are, and then keep deciding that for about seven years.

The AI Expo, and the question underneath it

The Atlanta AI Expo was where the abstract conversation got concrete. The framing across the floor was remarkably consistent: the interesting question in 2026 is no longer whether AI writes some of your code, but what happens to everything around the code once it does. Review. Testing. Documentation. The supply chain.

The thing we did not expect was what teams have started building for themselves. Ask someone what they shipped this quarter and a surprising number of answers were not customer-facing at all — an internal review assistant that knows the team’s architecture decisions, a deployment agent wired into their own runbooks, a support tool that reads the company’s real escalation history.

And when we asked how those tools connect to anything, the answer was almost always the same three letters. MCP has stopped being a novelty demo and started being infrastructure. Nobody was pitching it. It came up the way Docker used to come up in 2016 — not as the topic, just as the assumed substrate underneath whatever the topic actually was.

That shift has a consequence people are only starting to feel. Once MCP handles the plumbing, connection stops being the hard part. More than one person described the same failure: the connection works perfectly, and the agent still gives an answer their team abandoned two sprints ago. Beautifully plumbed, confidently wrong.

Hold that thought. We are coming back to it.

Robin Bender Ginn: Node.js runs AI, Who runs Node.js?

The OpenJS Foundation brought Node.js Interactive to RenderATL for the first time, and its Executive Director opened with the sharpest question of the conference. It was right there in the title: Node.js runs AI. Who runs Node.js?

OpenJS is the neutral home for 35 projects, including Appium, Electron, Jest, jQuery, webpack, and Node.js itself. That portfolio sits under an enormous share of what people are currently calling AI infrastructure. Every agent framework, every MCP server, every dev tool with a CLI is standing on some part of it.

Robin’s talk was about what it takes to keep that healthy: billions of downloads a year, a security surface that expands every time the ecosystem grows, and a maintainer base doing critical work on evenings and goodwill. She walked through what OpenJS has learned about long-term sustainability, and what concrete steps teams can take for the projects they already depend on.

The part that landed hardest was the asymmetry. AI-driven development has massively increased the consumption of open source — more code, generated faster, pulling in more dependencies — without a matching increase in resources flowing back the other way. Every company in that room benefits from the arrangement. Very few have a line item for it.

She was not asking for sympathy. She was asking for a plan.

Claudio Wunder: replacing what works, in Markdown

Then came the talk we will be citing for the rest of the year.

Claudio Wunder is a Platform Engineer at HubSpot, a Node.js Core Collaborator, co-chair of webpack’s Technical Steering Committee, and a voting member of the OpenJS Foundation. He shared the stage with Brian Muenzenmeyer — Node.js maintainer and author of Approachable Open Source — to tell the story of doc-kit.

The short version: the pipeline generating the Node.js API documentation has been running since the v0.6 era. It works. It has worked for more than a decade — long enough that if it were a person, it would be applying to middle school. And it needed to be replaced.

doc-kit is the standalone tool they built to do it. It parses, lints, and transforms Markdown into a stack of outputs:

• redesigned web pages
• legacy HTML, so nothing depending on the old shape breaks
• man pages
• JSON schemas
• search indexes
• llms.txt, for AI tools

Sit with that list for a second. One source format, written by humans, in a plain text file, feeding documentation for humans, documentation for legacy systems, structured data for machines, and context for language models. When the LLM requirement showed up, the Markdown did not have to change to pick up the job. It was already the right shape.

The part of the talk we keep thinking about was not technical at all. Claudio and Brian were direct about the real difficulty: refactoring something that has served the project since v0.6 while preserving trust. Contributors know the old pipeline. Users know the old output. Documentation is where a project’s credibility lives, and you get one shot at moving it without making people feel like the ground shifted underneath them.

“Replacing what works” is a genuinely uncomfortable engineering problem, and it is the one nobody puts in a conference abstract. They put it in the title.

Ethan Arrowood: Code and Learn, and testing that does not lie

Ethan Arrowood is Head of Open Source Engineering at Harper, a Node.js core contributor, and an active member of OpenJS and WinterTC. Also, and we mention this because it is a genuinely excellent detail, a professional ski instructor — which makes him the only person we met in Atlanta in August with a credible reason to want snow.

His session was on integration testing, and the framework he built at Harper to solve a problem most teams quietly work around: how do you test a complex system without simplifying it into something that no longer resembles production? Parallelizing tests when each needs its own process, dynamic port allocation so those processes stop colliding, and making all of it behave the same on a laptop and in CI.

It pairs directly with the AI conversation happening everywhere else in the building. If AI helps you write more code, your tests are the thing standing between “faster” and “faster at shipping bugs.” Tests that only pass because you simplified the system are worse than no tests, because they are confident.

Code and Learn

The other thing Ethan did at RenderATL was run Node.js Code and Learn, presented by RenderATL, the OpenJS Foundation, and Harper.

If you have not been to one: you sit down, get paired with a Node.js maintainer, and make a real contribution to Node.js in an afternoon. Not a curated toy issue. An actual one, in the actual repo, reviewed by the actual people.

The barrier to open source has always been less about skill and more about not knowing where to start or who to ask, and this format deletes both problems in a single sitting. It is also a direct, practical answer to the question Robin asked that morning. Who runs Node.js? Well: as of Thursday afternoon, a few more people than on Wednesday.

Markdown is the language agents run on

Put those four in order and something emerges. Justin argued that where you build matters. Robin argued that the infrastructure underneath all of it is maintained by humans who need real support. Claudio showed what that maintenance looks like done carefully, with Markdown at the center. Ethan turned a conference room into new contributors.

Now pick that thought back up from the expo floor. Teams are building their own software again, and MCP has become the wiring that lets it reach anything. That solves connection. It says nothing about knowledge — your agent can talk to fourteen systems and still not know your team abandoned the old migration pattern in March.

Connection is a solved problem now. Context is the open one. And Markdown is already sitting in that gap.

Open any modern agent setup and inventory the files that shape its behavior. SKILL.md. AGENTS.md. CLAUDE.md. The rules files, the runbooks, the architecture decision records. llms.txt, which doc-kit is now generating alongside the man pages. None of that is code. All of it is Markdown.

Somewhere in the last eighteen months, without a spec committee or a launch event, a format whose most advanced feature is the asterisk became the configuration language for AI systems. Not because it won a bake-off, but because when you need one format a model can parse reliably and a new engineer can read on their first day, plain text with light structure is the only thing satisfying both constraints at once.

Which raises a question we are going to spend the next few posts on: those skill files describe how your company deploys, escalates, and decides. That is proprietary knowledge in plain text, and most of it currently lives in whichever repo somebody dropped it in. If skills are IP, they deserve what any other IP gets — a controlled home, real permissions, and a record of what changed and when. More on that soon.

What we are taking home

The infrastructure conversation got serious: For years, “who maintains this” was a footnote. At RenderATL it was a keynote, delivered to exactly the people who needed to hear it.

MCP became infrastructure while nobody was announcing it: That half of the stack is largely settled. What an agent should know when it arrives is not.

Markdown is the language your agents run on: The format your team treated as documentation overhead is now shaping how your agents behave.

The South is not a satellite: Atlanta did not host a tech conference this year. Atlanta built one, filled it with its own artists, engineers, and organizers, and invited everyone else down to see how it is done.

If you were there, we would love to read your notes. Publish them, tag them, and they will show up in the HackMD Community where the next person searching for a RenderATL recap can find them.

And if you were not: 2027 is August 18 and 19. Start the conversation with your manager now — you have twelve months to build the business case, and roughly the same amount of time to acclimate to the humidity. 💜