AI Engineer World's Fair 2026

500 people vibe-coded for 30 days. I was one of them.

Loading slides…

Read the talk

500 people vibe-coded for 30 days. I was one of them.

Three projects trace Sanja Grbic’s shift from product designer to design engineer, showing how training, shared context and decision ownership changed the work.

From a talk by Sanja Grbic

Why speed looks different inside Automattic

What happens when a large software company gives small teams a month of autonomy to build and ship? Automattic’s Radical Speed Month organized work around autonomous two-person teams. Product designer Sanja Grbic built three projects during the initiative, each changing how she approached the next.

Automattic maintains WordPress.com, Jetpack and WooCommerce, alongside brands including Tumblr and Beeper. Grbic describes a company of approximately 1,400 people working asynchronously across a fully distributed organization. That working style had left a valuable foundation: years of documented decisions, concepts and discussions.

Grbic brought more than a decade of software design experience, including five years on Automattic’s Jetpack design team. She had watched both product interfaces and the teams behind them evolve. Design roles had expanded through web, UX, interaction and product design; tools had moved through Adobe, Sketch and Figma. AI added something different to that sequence: a tool that people across disciplines could use, with the same expectation of greater speed.

But speed depends on the environment. A team of one to three people building a new product can change direction with little coordination. An organization with thousands of people maintaining a product ecosystem built over five to twenty years has existing systems, dependencies and working habits to navigate. Faster implementation alone does not remove those constraints.

0:140:37
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:14 · section reference included

Make contributing possible before asking for speed

Automattic’s preparation began with access to AI tools and courses, then expanded into an AI enablement program. Grbic describes a two-week, role-specific course for every employee, combining lectures with focused hands-on practice. As coding tools improved, employees were encouraged to become comfortable contributing code, regardless of their original role.

That invitation depended on practical infrastructure. The systems and operations team provided security practices, processes and documentation clear enough for a non-engineer to start a working development environment. Access to a coding tool was paired with access to a usable development workflow.

For research and planning, Grbic also used Context A8C, an internal MCP server that gave AI tools access to Automattic’s accumulated knowledge and data. The company’s documented history could therefore inform new work through an authorized internal integration.

AI Enablement at Automattic slide listing tools, courses and two-week internal training; security and documentation; and Contex-A8C access to Slack, Linear and P2, beside outlined cubes.Open full source frame
Automattic’s AI enablement: training, security and documentation, and Contex-A8C.
3:073:20
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:07 · section reference included

Pause the roadmap and ship something real

Radical Speed Month made the change in priorities explicit:

  1. Pause roadmap work for thirty days.
  2. Pair with a partner.
  3. Ship something real.

AI use was welcome but optional. Participation happened in stages so the whole company did not stop its normal work simultaneously; roughly one-third joined the first round.

Grbic reports that approximately 500 participants started 794 projects during the thirty days. These were projects started, not a tally of shipped products. Her own month overlapped with the two-week in-person AI enablement course, which made AI tools central to what she could deliver.

Slide titled “Radical Speed Month” with three figures: 30 days, 501 people participating, and 794 projects started.Open full source frame
Radical Speed Month: 30 days, 501 participants, 794 projects started.

She was not starting from complete unfamiliarity with development: she had reviewed GitHub pull requests, tried Claude Code and built small hobby projects. But she had never built and shipped code at work. The experiment would take her from understanding development at a distance to owning parts of the delivery process herself.

0:144:21
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:21 · section reference included

A board game app becomes a lesson in collaboration

The first project began during training, with a concrete need: a board game evening was scheduled for that night. Two designers who were not comfortable in code, an engineer and a product lead had two hours to build something together. They chose a board game session manager.

Grbic reports that the four-person group built the app in two hours. Participants could create, view, join and manage sessions, start and stop games, see available games and seats, and chat. A 16-bit illustration of their actual office gave the app a recognizable setting that colleagues particularly liked.

The group agreed from the start that everyone would contribute. The engineer set up the GitHub project, taught basic commands and explained how Git worked. Grbic had used versioning before and thought she understood Git; working on a shared repository exposed the gap between familiarity and being able to use it confidently.

Sitting together, the group divided the work, tested the app and committed code. Because the project was separate from Automattic’s existing products, they could concentrate on learning without putting those products at risk. Development happened in Claude Code; Grbic recalls Nano Banana as the only additional visual tool, used to create the office illustration.

An engineer’s impact can grow by enabling other people to build. The engineer had not simply taken the implementation away from less experienced teammates: he had helped them participate. Grbic extends that lesson to design and product work, too. As AI makes it possible to contribute outside a primary discipline, the person with the strongest relevant skill can help the whole group improve.

5:596:07
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:59 · section reference included

Build the tracker, then own its delivery

The next project put that new confidence to work on an organizational problem. Grbic joined an engineer and a design operations colleague to make relevant, current design system information easier for humans and AI tools to find. Their open-source design system served products inside and outside the WordPress ecosystem and changed continually. The problem was broad enough that each collaborator took a different part.

Grbic proposed a design system status tracker. The engineer raised questions about performance, maintenance and whether the idea could be built at all. Encouraged by the training exercise and allowed to experiment, she decided to implement the full idea as a proof of concept herself. Discovery involved three people; design and implementation became her responsibility.

She built it piece by piece. Getting live component previews working was a turning point: the tracker could show the components themselves, rather than only point to documentation. It gathered links from GitHub, Storybook and Figma, sorted them by status and tagged them with the appropriate library names. Grbic reports that search worked well; she shared the tool with fellow designers and incorporated their feedback.

Design System Tracker with text in its search field and a single Button card showing Save, Cancel and Reset controls.Open full source frame
The Design System Tracker displays a Button component as a filtered search result.

The working prototype came before Figma. When she could not communicate a visual refinement effectively to Claude Code in words, she used Figma to create an image that conveyed it. Figma became a tool for resolving particular visual questions, while the live project remained the main place where she iterated.

Grbic reports building the prototype within one week, followed by another week and a half for fixes, hosting and deployment. That second period included rebuilding parts of the implementation and delivering two versions:

  • Internal tool: deployed on Automattic’s internal platform.
  • Public-facing site: published with a curated set of data and information.

A working prototype was therefore only part of the job; delivery required deciding what each audience could access and getting the tool hosted.

For Grbic, delivery marked the move from designer to design engineer. AI helped, but she credits ownership of the entire process most of all. Radical Speed Month gave her permission to experiment and make the necessary decisions, reducing time spent negotiating and handing work between roles. She could move directly from identifying a need to implementing and deploying a response.

9:359:51
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:35 · section reference included

Two designers build a merchant chat proof of concept

With roughly a week left, Grbic paired with another designer. Grbic reports that they built a working iOS chat proof of concept for WooCommerce merchants in six days. Merchants could answer shoppers in real time from their phones or enable AI to respond.

The proof of concept connected several parts of the existing ecosystem:

  • Merchant access: authentication through WordPress.com and a Jetpack connection.
  • Visitor interface: a chat widget that inherited the website’s theme.
  • AI responses: an agent that scanned the site for information and answered visitors’ questions in real time.

Grbic also describes the agent as distinguishing questions it could answer from those it could not. The delivered artifact was a working proof of concept; the account does not establish a production launch or measured answerability accuracy.

13:2313:36
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

13:23 · section reference included

Keep shared context beside the prototype

The two designers found it straightforward to align on features because both approached the work through users’ needs. They began exploration in a Claude Code project folder, recording their chats and ideas in the file system. Grbic found that this practice sped up collaboration and subsequent building enough that she continued using it in daily work. The project folder held the thinking behind the implementation as well as the work itself.

The main design artifact changed, too. Previously, the team would stay in Figma until the design reached high fidelity. Here, they developed a plan, built the working prototype and then returned to Figma for a mood board and UI refinement.

StagePrevious workflowWorkflow used for merchant chat
Before implementationDevelop high-fidelity designs in FigmaEstablish a plan
Main iteration artifactFigma designWorking prototype
Visual refinementBefore buildingReturn to Figma after building

The sequence let them work on functioning behavior earlier while retaining Figma for visual decisions it handled well.

14:3914:55
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:39 · section reference included

Give people room to change how they work

Over the month, Grbic changed a process she had followed for years. The first project showed how an engineer could become an enabler and teacher as more colleagues began contributing code. The second expanded her own responsibility into design engineering: pushing code to production within a large, established system that required substantial onboarding and security practices. The company’s initiatives gave her room to find a new workflow within those existing requirements.

By the third project, the change was shared. Both designers had learned how to apply the tools and could work from a common understanding. Grbic attributes their speed to that mutual adjustment, not simply to one person becoming more proficient. The same kind of experiment could begin within a team rather than waiting for an entire organization to change.

Changing tools and processes requires changing the human behavior around them. Give people access to the tools, identify colleagues who can teach and champion their use, create space for experimentation, and grant enough agency to make decisions. Those conditions let people break established habits and discover where AI can actually make their work faster.

Slide stating that changing processes or tools in large organizations requires shifting human behavior, above four circles labeled Executive support, Change champions, Experimentation and Agency.Open full source frame
Four catalysts of change: executive support, change champions, experimentation and agency.
16:0016:21
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

16:00 · section reference included

Resources

From the talk

Transcript download
Full transcript

Loading transcript…