AI Engineer World's Fair 2026
How a Remote Company Builds AI Fluency — Em Shreve, Automattic
Read the talk
How a Remote Company Builds AI Fluency
Em Shreve explains how Automattic combines two weeks of protected learning and real project work with shared tooling, peer teaching and follow-up support across a fully distributed company.
From a talk by Em Shreve
At a glance
Ideas worth remembering
Automattic protects two weeks for learning, split equally between facilitated workshops and real project work, with peer demos throughout.
The engineering curriculum progresses from using skills and MCPs to creating integrations, extending agent harnesses and running autonomous routines; it requires continual updates.
Graduates extend the program through teaching and support, while shared MCP infrastructure and skills let their projects become reusable company capabilities.
Surveys at 72 hours, 30 days and 90 days track continued development, alongside intentionality, workflow fit, sharing and comfort with specific capabilities.
Cohort projects included a Zendesk connector adopted by multiple teams, a cross-tool skills directory and a guided support tool built by a support teammate.
Tool access does not create time to learn
A new AI tool can be available to everyone while everyday work barely changes. Em Shreve, Automattic’s head of AI enablement, opens with that practical problem. The company behind WordPress.com and Tumblr has over 1,400 people in 82 countries. Its bet is that better work requires giving those people real time to learn and experiment.
For this fully distributed company, the intervention is deliberately in person: a two-week immersive enablement program. Half the time goes to facilitated learning, organized by the learning and development and AI enablement teams, with colleagues hosting workshops and hands-on activities. The other half goes to real project work. Participants bring something they need to finish, or build a tool or dashboard that would help in their daily work.
Projects can also produce reusable skills or MCP integrations that remain useful after the two weeks. Peer sharing makes these experiments visible: participants show their workflows and give project demos at the end of each day. The learning environment therefore includes both guided practice and a view of what colleagues are managing to build.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Move from using tools to extending them
The program has three forms, each serving a different kind of work:
- Engineering cohorts: Go deeper into applied AI and the machinery engineers can extend.
- Cross-functional cohorts: Bring product, engineering and design into the same room to explore how their work together changes.
- Non-technical workshops: Extend learning to finance, executive leadership and other functions.
The engineering curriculum starts with foundations: creating skills, using MCPs, spec-driven development and responsible AI use. It then moves from consuming existing capabilities to creating new ones. Participants learn to create MCPs for new functionality and use the hook systems of Codex and Claude to extend the agent harness—the surrounding system through which an agent works. Later topics include autonomous agents and routines that run in the background.
The final applied workshops connect those capabilities to familiar work: WordPress AI features and building a personal assistant. The progression matters. Someone begins by learning how to use an existing tool, then learns how to add capabilities, and finally applies that knowledge to something they can use themselves.
This curriculum carries an ongoing maintenance cost. Model releases and tools change quickly enough that writing the course once would leave later cohorts learning stale practices. Automattic continually updates the material so a new cohort does not receive instruction that is already six months out of date.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Carry the learning back into a distributed company
The second form of compounding is teaching. At the time of the talk, Automattic had run seven cohorts of about 45 people each. Its year-end target was around 600 participants, reaching about two-thirds of its engineers as well as people in other functions. Each graduate is intended to become a “multiplier” who helps colleagues the central team has not yet reached.
Several mechanisms carry that knowledge beyond the original room:
- Guides: A developing program recruits strong practitioners to help spread knowledge into other parts of the company.
- Local AI get-togethers: Graduates organize regional events and workshops with support from the enablement team. About ten had taken place around the world.
- AI ride-alongs: Short workshops show how someone actually works with a tool. Watching a colleague in a similar role makes the workflow easier to grasp than hearing only the concepts.
Shared skills and MCP infrastructure let colleagues build on each other’s work. An always-on Slack help channel provides a place for newer users to ask questions at any point in the day, with program graduates asked to help. These channels serve different needs: a ride-along makes a workflow visible, a shared skill makes work reusable, and the help channel gives someone a way forward when they get stuck.
How does a two-week gathering reach people who never attended it? The diagram separates the two routes out of a cohort: graduates teach colleagues, while their projects add reusable capabilities. Both extend the reach of the original investment in learning.
Facilitated learning, real project work and peer demos.
People carry working knowledge through teaching and support; projects carry capabilities through shared infrastructure.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Measure how AI fits the work
The program tracks changes in how people work, including three distinct dimensions:
- Intentionality: Choosing thoughtfully when and how to apply AI, including moving beyond chatbot-only use where an agentic workflow is useful.
- Flow: Integrating AI into a workflow in a way that makes sense for the person doing the work, rather than using it for its own sake.
- Sharing: Helping knowledge reach other people in the company.
Individual measures also cover skills and custom commands, comfort using and applying MCPs, and context management. Follow-up surveys run 72 hours, 30 days and 90 days after the program. Shreve reports continued growth in confidence and experience during those 90 days as participants have more time to experiment and apply what they learned. These are survey-based observations; they do not establish the size of a productivity gain or isolate the program’s causal effect.
The follow-up schedule treats the two weeks as the beginning of a change in working habits. Participants leave with capabilities to try, then learn more through use. Measuring again after 30 and 90 days captures that continued development instead of treating confidence immediately after training as the final result.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
What two weeks of real project work produced
The closing examples make the program’s output concrete. One group built an agentic analytics system for WooCommerce, WordPress’s e-commerce plugin, so users could ask questions about their store and receive answers. Another early cohort wanted to use AI tools for support and built a Zendesk connector on the shared internal infrastructure.
The Zendesk project shows the full path from a local need to a company capability. A cohort identified support work it wanted the tools to handle, added a connector, and made that addition available through shared infrastructure. Various teams then adopted it. The observable change is reuse across teams: work done during one cohort became something other groups could use afterward. Other providers were added where native MCP integrations were unavailable or where Automattic wanted additional observability.
A cohort also produced the company’s agent skills directory. Automattic does not prescribe a single tool: people can use Codex, Claude, OpenCode and others. That choice makes cross-tool reuse a requirement for the directory. Shared knowledge needs to remain useful across the tools employees choose, rather than being confined to one agent environment.
The final example is guided admin support. During a help session, a support teammate wanted to show someone where something was in a dashboard. The resulting tool could guide the session and highlight parts of the screen. Its builder was a support colleague who felt more comfortable after the program. The example expands what project work can mean: someone close to a recurring support problem gained enough confidence to build a tool for it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Make room for experimentation
Shreve’s closing instruction is to “give people real time to actually experiment and learn.” Automattic’s implementation makes that time concrete: protected weeks, facilitated workshops, useful projects, peer demonstrations and help after participants return to work. An education program has to make the learning environment central if it expects people’s way of working to change.
The talk ends by pointing to WordPress.com and WooCommerce’s official AI connections and MCPs as further areas for discussion. Those product integrations sit alongside the internal enablement effort: employees are learning to work with AI while also creating and extending capabilities that others can use.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Read the complete timestamped transcript
- 0:12
All right. Hi, everybody. Can you hear me? Cool. All right, so, uh, my name is Em Shreve, and I work at Automattic. Um, and if you're not aware of what Automattic is, we are the company behind WordPress.com and Tumblr and a number of other properties. Um, and we're a fully distributed company, uh, with over fourteen hundred, uh, people in eighty-two countries. Um, and I'm here today to
- 0:42
talk about our approach to, uh, AI enablement and upskilling our organization, um, to use all of these, uh, new tools and ways of working. Um, so our bet is that you don't just get good quality learning or work out of just handing people the tools. You need to give them real time to learn and,
- 1:12
uh, experiment. Um, so what we've been doing is running these two-week, um, enablement programs where we bring people in in person for a two-week immersive experience. And, uh, the way that we do that is, um, through fifty percent facilitated learning, um, through our, uh, learning and development team and our AI enablement team. We pull together workshop hosts
- 1:42
from within the company to run workshops and hands-on, um, activities. Uh, and we'll look a bit at what that curriculum looks like shortly. Um, and then fifty percent real project work. So we have folks bring in, um, real work that they want to get done or a tool or dashboard that would help them in their day-to-day work. Um, or these can also be creating new skills or
- 2:12
MCPs to enable them to work faster outside of the program. And another large part of this is, uh, encouraging peer sharing. Um, so sharing their own workflows to show, um, kind of the art of what's possible. And, uh, as they work throughout the two weeks, we also do a number of, uh, project demos at the end of each day.
- 2:41
So we've, uh, we've ran three flavors of this. Um, our main set being for engineers, where we go, uh, a lot more in depth, um, getting into applied AI. We've also, um, experimented with cross-functional programs, um, with product engineering and design getting together in a room and discussing kind of the future of, of how we all work together, um, with, with these new tools. And then we've
- 3:11
also extended into non-technical teams, and run workshops with our finance group, our executive leadership team, um, and other functions as well.
- 3:26
So our engineering curriculum starts the first few days, uh, covering the basics, skills and how to create them, um, learning how to use MCPs and spec-driven development, um, as well as kind of the foundations of how to use AI responsibly and how to, um, apply it in a way that's mindful. Um, and then we start getting into deeper topics, uh, creating MCPs,
- 3:56
um, for new functionality, getting into the hook system of Codex and Claude, and being able to extend the harness that way, um, up to, uh, autonomous agents and running routines and such in the background. Um, and then we end on some different applied topics. Um, so a-again, Automattic is behind WordPress.com. So, uh, we normally run a workshop on applied WordPress
- 4:26
AI, um, and some of the cool features that are in WordPress core, uh, around AI, as well as, uh, being able to create your own personal assistant. So we really try to walk people, um, from the foundations up into, uh, the ability to start creating these new things for themselves.
- 4:49
And one thing that we found out is you cannot just write this curriculum once. As we're all aware, AI is constantly changing from day to day, week to week. The top model might be different, uh, depending on the latest release. Um, so we're constantly needing to update the curriculum and stay current to make sure when we are running these programs, we're teaching the right thing and we're not teaching six months out of date.
- 5:17
Um, another interesting part I wanted to highlight here is we have our own, uh, internal, uh, MCP server called Context Automattic, and this allows us, um, to rapidly add new tooling sets for MCPs that might not exist or that we want to build our own, um, like, governance and observability layers on. Um, and one of the cool things that, uh, I'll highlight shortly
- 5:47
is, uh, some of the projects coming out of these cohorts is extending these abilities. So as people go through, the capabilities, um, compound, and we're able to do more as a company, uh, uh, because of that.
- 6:06
Another main focus of our program is that these aren't just a number of people that we're sending through. We're also wanting each person we send through, um, to be a multiplier, uh, within the company and extend our reach to the folks that we can't get to yet. Um, so, so far this year, we've ran seven of these cohorts of groups of about forty-five people. Um,
- 6:36
and we're aiming to get around six hundred people through these programs by the end of the year, um, which is, uh, about two-thirds of our engineers, as well as, uh, some of those other focus areas that I highlighted at the beginning. And we would like to equip t- all these people with, uh, ways of sharing their knowledge out, and again, extending our reach. So we've tackled this in a few different ways.
- 7:07
Um, we are beginning a guides program, or, um, I think other companies call this, like, their champions program. And what we want to do is extend our team's reach into other areas and have these, um, like, top practitioners that we've identified within the company be able to help us out and, uh, continue spreading the knowledge. Um, because we're distributed,
- 7:37
we also have began these local AI get-togethers. So people that have gone through the program, um, are able to then run these, like, little events within their regions and, um, run workshops, and we'll help support that as well. Um, and we've done about ten of those so far in various different areas of the world. Um, and it's been, been exciting to see, like, what some of these people that have gone through the program have,
- 8:07
have gone on to do and teach. Uh, we also, um, organize these things called AI ride-alongs, and these are little workshops that we run, um, basically showing how you work, how you use a different tool. Um, because what we found is seeing someone in your similar role or, um, in, in a similar job to you, being able to see how they use it, uh, is a lot more, um, easy to
- 8:37
grasp than just, you know, talking about the concepts. Um, and then we're also attempting to build a shared skill and MCP infrastructure. So again, that context agency, um, system that I mentioned, but also, uh, encouraging people to create and share skills and build, uh, on top of each other's work so that we're all kind of building, um, towards the same
- 9:06
kind of platform. And then we have an always-on, uh, help Slack channel where we also ask people that have gone through this program to help out, um, and just kind of be available for anyone newer, uh, to working this way to have a place to come and ask questions at any point of the day.
- 9:31
So I'm gonna share a few of the, the metrics that we've been tracking and, um, improvement we've seen throughout the program. Um, one of these is intentionality and, um, just applying to-- applying AI to your work in a more thoughtful way. Maybe these people came in at the beginning just using it as a chatbot or, um, you know, not fully using an agentic workflow.
- 10:01
Um, so having people be a bit more intentional about how and when they apply AI. Um, flow, which is, like, how well they're able to integrate AI into their workflows. Um, if f- people feel as if, um, they're able to use it in a way that makes sense for them and not just using it for AI's sake. Um, and then the, the sharing part again is very important to us so that we can continue, um,
- 10:31
reaching other people in the company and, um, kind of making everyone, uh, a, a teacher and an expert on these topics. Um, we also track, uh, individual lift on various metrics such as, um, skills and custom commands, um, how comfortable they feel using and applying MCPs and, uh, other topics like context management. Um,
- 11:03
and what we've seen, uh, from doing various surveys, uh, after the program, so we'll do one seventy-two hours after, uh, thirty days after, and ninety days after, people will continue growing within those ninety days and, um, being able to apply things more and, like, they've h- had more time to experiment and try things. Um, so we're still seeing people, like, gain confidence
- 11:33
and experience, uh, even once they've gone through the program.
- 11:42
Um, so I just wanted to highlight just a few of the things that folks have built, um, just within the two weeks of the program. Um, one of these is, uh, an agentic analytics system. So, uh, we ha-- WordPress has a, uh, e-commerce platform plugin called WooCommerce. And some of the people that went through our program, um, created a
- 12:12
agentic system for asking questions about your store and getting back answers. Um, a number of folks that go through, uh, build on top of Context A and C. Um, so one of our earlier cohorts, um, they wanted to be able to use, uh, some of these tools for support, um, and built a Zendesk connector, um, which now we've had adopted through various teams throughout the company. Um, we've built other providers in there
- 12:42
that, uh, might not have a native MCP or, again, because we wanted to add additional observability there. Um, we've been building out our own agent skills directory. Uh, we don't prescribe a single tool, so folks are able to use Codex or Claude or, uh, any other number of tools, OpenCode, et cetera. Um, so we wanted to build our skills directory in a way that can be used across all of
- 13:12
those. Um, and the whole skills directory actually came out, out as a project from one of these cohorts as well. Um, and then finally, another one I wanted to highlight here is, uh, guided admin support. So, um, during a help session, say you wanna show someone exactly where something is, uh, within the dashboard, um, they built a tool using, uh, Cloud
- 13:42
Code that allowed them to be able to, um, guide and lead a session, highlight different parts of the screen. Um, and this actually came not from an engineer, but one of our support folks, um, that felt a bit more comfortable after going through one of our programs. And
- 14:04
that is the end of my presentation, but the one takeaway that I wanted people to take from this is, uh, give people real time to actually experiment and learn. Um, you can't just throw a bunch of tools at them and expect their way of working to change. You need to, to make space and give an environment for learning, um, and, and make that, like, the star of whatever your education program is.
- 14:34
And, uh, I will be available for questions at our booth, which is, uh, the Automattic booth right over on the other side near that stage and the swag pickup. Um, so I'll be there for a little bit after this talk if you wanna talk about anything I talked about here, um, or you'd like to hear more about how, uh, W- how WordPress.com and WooCommerce are, uh, embracing
- 15:04
AI, um, with our official connections and MCPs. Um, so please stop by the booth. Um, and yeah, thank you.