Jokūbas Leikauskas, Editor, Nucleic Acid Insights, speaks with Edward Miracco, Head of Enterprise AI Enablement, Moderna, about enterprise AI adoption at Moderna, AI-driven lipid nanoparticle and neoantigen design, rare disease applications of mRNA, AI in regulated and clinical settings, and the future of RNA therapeutics.

What are you currently working on?
EMMy role at Moderna is to lead the Enterprise AI Enablement team, which has evolved significantly over the past 6–9 months. Initially, the focus was on how to roll out AI tools to everybody to ‘vibe code.’ I like to describe the large language model (LLM) AI revolution at Moderna in three epochs.
Pre-December 2025, the teams in the digital organization started to use LLMs for code generation and think about how to build the infrastructure to let everyone at the company access LLMs.
We signed an enterprise agreement with OpenAI a few years ago – we were one of their first full enterprise partners. This gave the whole company the opportunity to leverage ChatGPT in their work, and a lot of the Microsoft ecosystem connectors have been valuable. The OpenAI – Microsoft partnership and experience at Moderna has been valuable.
Then, over the Christmas and New Year week of 2025, a lot of people were off and had free time, and all of a sudden everyone started realizing that Claude Opus 4.5 was good at coding. That spark lit the ‘vibe code’ epoch. We all came back from the holiday break, and within a week or two, our CEO and the executive team realized that we could start solving a lot of our own business problems by putting this tool in the hands of the business experts. There was a Cambrian explosion of vibe-coded apps, and everyone was excited, but the excitement was short-lived.
The third epoch arrived around March 2026, with the realization that vibe-coded apps are not enterprise production apps. To graduate from a vibe-coded app into a production app, you needed digitally resourced experts, software engineering experts, to say, for example, ‘You built this with a Node.js backend, but we want to use a different one,’ or ‘You are using DuckDB, but we use MySQL.’ Those mismatches of infrastructure were the big issue, but cybersecurity was a quick runner-up. The cybersecurity issues taught us quickly that if we are going to put this tool in the hands of business experts, we also need to provide the guardrails to make sure they can do it safely. That led us to build out what we are calling a work reinvention team, or work reinvention group. We have selected up-and-coming AI bright stars across every enterprise department and given them training on agentic engineering tools.
We are an OpenAI shop, so we use Codex, but this is specifically in Visual Studio Code, and we feel strongly that that is a winning harness and LLM connection. From there, we designed tight AGENTS.md files, then built several infrastructure improvements for skill, centralized management, as well as connectors and Model Context Protocol governance.
I am a scientist by training. I spent many years at Moderna in the science side of the business, innovating and inventing the whole way through. I am a ‘move fast, break things’ kind of thinker, but this role is not that. This is an enterprise-facing role, and I have had to realize that you cannot just give a powerful tool to a thousand people and expect nothing to go wrong.
As we matured, we had the guardrails, we were starting to build the governance, and we knew what these tools could do. The realization has been that when you put the tool in the hands of the experts, all of a sudden, they are able to build a prototype to show their digital partner what they need. That has shifted the build versus buy decision more toward build, which means you also need a significant number of software engineers on staff to do that professional building.
What happened from that realization was two things specifically. First, you can build new deterministic software with agentic engineering, with LLM coding, and that final product may not have any AI in it at all. That is one acceleration that has happened from coding LLMs. The second is building a deterministic harness or a framework around active use of LLMs in that piece of software’s workflow. That is active AI being used to do something that previously you could not do.
We needed a place for these two types of products to live, because hosting them on random servers is not a platform. We built an internal platform called Workbench, and a little sibling to it called Workbench Lab. If you have built an application that solves a significant problem, and it is so valuable that the company would be impacted if it went down, it needs to go to Workbench. Applications in Workbench require a digital representative, PagerDuty, and cyber review. A smaller app that helps a handful of people, and that would inconvenience your team but not impact the company if it went down, can sit in Workbench Lab. There, it is limited to smaller numbers of people and we still have cybersecurity involvement, but the blast radius is small, and it is kept that way.
The scientists have started building bespoke analysis software following the rollout of these tools, and that has accelerated our number of programs. For example, Moderna has to make COVID-19 and flu vaccines every year. That pipeline involves many mRNAs, all around the world, and you need to be able to refresh new sequences as fast as possible, because it is a reproducible flow.
We have looked at the process end-to-end and then used AI to improve it through a collection of agent skills, while also building deterministic software along the way that fits into the workflow that Moderna specifically does. This is around 6,000 steps, from a new COVID-19 variant genome being sequenced to shipping the product to a pharmacy.
Across every one of those steps, it is hard for a human to keep it all straight in their head, so we use AI significantly to help us map it all, build the solutions, and figure out which workflows we need. That has been a major value-add for speeding up and reducing the amount of toil that our scientists, engineers, and businesspeople have to do.
I would say the biggest positive impact in our space has been in the reduction of toil across the board for digital or computational things.