Project AEGIS- Why we’re accelerating our AI Journey

Like almost every other tech company, ManoMano’s engineers have embraced AI copilots. They are fantastic tools. They complete our thoughts, speed up boilerplate, and act as great rubber ducks.

But a copilot has a fundamental limitation: it only works when you are working. It sits in the passenger seat, waiting for you to turn the wheel. And without context, copilot is a generic developer, not a ManoMano developer.

At ManoMano, we are DIY enthusiasts. When we see a limitation, we build a tool to fix it, “A la Mano” way. We started asking ourselves: How can we accelerate delivery? What is needed to leverage AI in everyone day-to-day work? What happens when we move the AI from the passenger seat to the driver’s seat? Can we build a squad of semi-autonomous agents that handle maintenance, documentation, and analysis while our human engineers sleep, so they can focus on pure innovation during the day?

This isn’t Sci-Fi anymore. It’s an engineering challenge. And we decided to tackle it.

Providing AI Tools to the Engineering Teams

Dev Life

Even if you learn at school that a developer’s job is to produce code, reality is far from that. From updating JIRA and migrate legacy code to analyzing incidents, dev life is way more complex and sometimes full of meetings. That’s why time spent delivering value is rarely close to 100%.

We aren’t trying to replace developers. We are trying to replace toil tasks (and not core work). The idea is to give our tech team superpowers by offloading the tasks that drain time and energy for no added value.

It’s already the case for dependency update, for example. We introduced a while ago Renovate Bot who is checking every night the outdated dependencies for each project, try to update them and propose a Pull Request if CI is green. If you don’t know this project, it’s worth taking a look.

So we are trying to identify these chore tasks and will try to give developers a way to do it faster, keeping the same control and quality. It could be by providing knowledge, feedbacks, tools and services.

Some ideas: ticket refinement, bug reproduction/analysis, migration following a migration guide, expose a controller into a new technology (GraphQL), …

Semi Autonomous Agents working during the night

The Team

To build this, we didn’t create a massive committee. We formed a “Two-Pizza Team” of 8 specialists — Architect, SRE, Data Engineer and Senior Developers (covering all languages). It’s not a massive R&D department, it’s a focused agile team. The idea is to iterate weekly, move fast and act autonomously.

AEGIS was born as ManoMano’s internal program to scale AI-assisted development and introduce human-supervised autonomous agents for repetitive engineering tasks.

Our mission is simple but highly important: build ManoMano’s next generation of AI-assisted engineering. We decided to split the work in two phases:

• Onboard everyone to Dev Assisted by AI, improve usage and share knowledge. It will be challenging to convince everyone, but we believe stories and real success will prove AI is a new great tool.
• Thinking, designing and developing semi-autonomous agents. These, triggered automatically or on demand, will be able to achieve chore small tasks. The idea is to define a Human in the loop strategy to decouple workforce by keeping control and refreshing developer experience.

Boundaries

Three main topics are very important in this adventure.

Cost - How much a developer can spend to automate tasks during a month ? Should we target a fixed amount ? How can we optimize LLM usage ? We will need to put in place monitoring (at team level) and measure Lead Time to Change.

Security - What LLM to use (internal, European, leader) and how to ensure produced content doesn’t put ManoMano at risk ? We need to put clear boundaries and controls to ensure we are not spreading our code everywhere, and we do not introduce security flaws.

Quality: What is the purpose of generating 1000 code lines per minute if you trash away 80% ? Our quality levels must be respected, and we will iterate to improve prompts, context, knowledge to obtain similar or better results than what we produce today.

Why this series of blog posts ?

We decided to document this journey openly. We will share our architectural decisions, our stack choices, our successes, and yes, the embarrassing moments when our agents completely hallucinate. Obviously, we will not be able to share everything, but you will follow the story (written by humans).

This is bleeding-edge territory. There are no established playbooks for deploying agents at scale in a large enterprise. We need to research, propose, listen to feedback, iterate and improve.

Writing memories while agents are working

What’s Next?

In the next post, we’ll open up the hood. We’ll share the first steps of AEGIS program. You will understand why we are designing for model agnosticism from Day 1 (sorry Claude, we need to keep our options open for Gemini), and how we are structuring our tech stack. See you in 1 month.