How Monzo built Agent Chip on its engineering platform

Felipe introduces Monzo as a digital bank trying to make money work for everyone.

Users can @-mention Agent Chip to get it to do work, answer a question, explain an architecture, or just chat.

It also performs code reviews and can be assigned Linear tickets, which it works on directly.

Agent Chip now authors over 10% of all merged PRs across tech teams, including data, ML, back-end and app engineering.

It runs more than 1,800 tasks per day, and business users have started using it for questions, queries and self-serve needs.

The talk is not about bragging over the agent's launch. It aims to share what made the product succeed.

The presenters say the 'LLMs' in their framing stands for lots of little multipliers: capabilities the platform already had.

Before any remote agent existed, Monzo began investing in supporting AI coding agents in local environments.

That groundwork came first, and it is what later let the remote agent connect to internal systems.

The first thing Monzo built was an MCP gateway giving easy access to knowledge sources and internal tools.

The gateway sits in front of internal services and decides what agents can access, while also managing authentication and authorization across many agents.

The guiding principle is that the more context and data an agent can safely get, the better.

Remote agents run inside sandboxes in an isolated Kubernetes namespace, and the orchestration system triggers them and reads their output.

Each sandbox has an agent image defining exact tools, CLIs and credentials obtained from the gateway. A proxy allows egress only where explicitly chosen, so code repositories are reachable on purpose.

The agent itself runs in the innermost box, where the non-deterministic behavior happens.

With the gateway and orchestration in place, the agent becomes a building block that later use cases can plug into.

It can connect to Slack, GitHub and Linear, as well as the more than 3,000 services in the platform.

Engineers can run a broader range and volume of tasks and delegate to remote agents instead of babysitting them, which frees people for ambiguous problems.

Monzo deliberately limits its chatter during incidents so it does not flood the channel while testing hypotheses. Humans control how much it contributes.

Ask channels are squad-to-squad interaction inside Monzo, where a team needs context on a back-end service, a product or the roadmap.

Because Agent Chip connects to knowledge sources, it drafts a reply for a human to review. Sometimes the draft solves the ask outright; otherwise it asks for more context and tags a person.

The proactive version monitors backend changes and estimates their likely effect on data before the next batch run.

Non-specialists can start from product feedback, ideas or discussions by calling Agent Chip on Slack and getting a draft already tested in staging.

Only then is an engineer involved to review and deploy it. The rigor of Monzo's software process does not change, but more people can contribute.

Francesco argues that the agent is now a new building block that enables new agentic use cases quickly, and that it connects safely to security and platform requirements.

The agent layer can be swapped for any model and chosen per purpose, so Monzo is not tied to a provider.

Francesco says Monzo is moving to cover a bigger development window: drafting back-end and app changes, testing them in staging, and covering data tests.

He adds that even if development stopped today the value would already be real, and that end-to-end remote agents would integrate the systems better.