# A Bank Built Its Own Coding Agent. It Took Off

Felipe Felisola Caso & Francesco Galletta, Monzo | AI Engineer NYC 2026 | 15:30

Source: https://www.youtube.com/watch?v=adfw3xJBx1o
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/a-bank-built-its-own-coding-agent-it-took-off
Published: 2026-10-20
Tags: agents, coding-agents, harness-engineering, mcp

## TL;DR
- Monzo's Agent Chip grew from a Slack reaction into an agent that authors over 10% of merged pull requests and runs more than 1,800 tasks each day.
- Its rapid adoption depended on existing platform capabilities, including an MCP gateway, isolated sandboxes, deployment systems, staging environments, and internal knowledge access.
- Monzo now uses Agent Chip for incidents, data quality alerts, service questions, and work proposed by non-engineers, while keeping human review before deployment.

## Summary
Felipe Felisola Caso and Francesco Galletta describe how Monzo built Agent Chip, an in-house agent that began as a Slack emoji reaction. Employees can ask it questions, review code, investigate incidents, and take on Linear tasks. It now authors over 10% of merged pull requests across Monzo's technology teams and runs more than 1,800 tasks a day. The speakers attribute this growth to many existing platform capabilities rather than to the model alone. Monzo built an MCP gateway for controlled access to internal tools and knowledge, then added orchestration that runs agents in isolated namespaces with restricted tools and network access. Agent Chip has become a reusable platform component for incident response, data quality monitoring, service questions, and work initiated by non-engineers. Monzo keeps engineers involved in reviewing and deploying changes. The system can also switch between model providers and may eventually cover more of the development process across backend, app, data, and machine learning systems.

## Key ideas
### Monzo's coding problem spans thousands of services and data systems
[00:01](https://www.youtube.com/watch?v=adfw3xJBx1o&t=1s)
Monzo is a digital bank with engineers maintaining over 3,000 backend services, around 13,000 data warehouse models, and dozens of machine learning predictors. These systems support decisions about credit and financial crime, along with the bank's app, backend, and data products. That scale gives a coding agent a large set of services, repositories, and internal knowledge sources to work with. It also creates strict requirements because changes can affect regulated banking operations.

### Agent Chip spread because employees could invoke it directly in Slack
[01:25](https://www.youtube.com/watch?v=adfw3xJBx1o&t=85s)
In January 2026, a Slack channel for Agent Chip went viral inside Monzo. People could react to a message with the agent's emoji and have it pick up the message as a prompt. The tool then became an always-available teammate in Slack and Monzo's wider environment. Employees could ask questions, investigate architecture, chat with it, request code reviews, or assign Linear tasks. The low-friction entry point helped the agent move beyond a specialist coding tool.

### Agent Chip reached substantial use across Monzo's technology teams
[02:50](https://www.youtube.com/watch?v=adfw3xJBx1o&t=170s)
Agent Chip authors over 10% of all merged pull requests across Monzo's technology teams, including data, machine learning, backend engineering, and app development. It also runs more than 1,800 tasks per day. Business users have started using it to answer questions, write queries, and handle other requests. The speakers present these figures as evidence of adoption across the company, rather than as a demonstration limited to one engineering group.

### Existing platform capabilities multiplied the value of the language models
[03:51](https://www.youtube.com/watch?v=adfw3xJBx1o&t=231s)
The speakers explain the acronym LLMs as 'lots of little multipliers'. Agent Chip benefited from capabilities Monzo already had, including alerting, monitoring, third-party integrations, staging environments, queuing abstractions, deployment systems, and database abstractions. Those systems made it easier to build, test, and operate the agent. Monzo had also started supporting local coding agents before Agent Chip became a remote agent, so the platform work accumulated before the product became widely used.

### The MCP gateway gives agents controlled access to internal knowledge and tools
[04:51](https://www.youtube.com/watch?v=adfw3xJBx1o&t=291s)
Monzo built an MCP gateway to connect coding agents to internal knowledge sources, tools, and services. The gateway sits between agents and the systems they need to use. Monzo chooses which capabilities to expose and manages authentication and authorization for different agents. The speakers say that giving agents more context and data is useful only when that access is provided safely. The gateway gives them a controlled way to provide that context.

### Sandboxed orchestration limits what a remote agent can do
[05:59](https://www.youtube.com/watch?v=adfw3xJBx1o&t=359s)
Agent Chip runs through an orchestration system in an isolated namespace. Multiple sandboxed pods can be created, and the agent image defines its tools, command-line interfaces, authentication, and authorization. A proxy controls access to external systems, including code repositories, and Monzo chooses the allowed network egress. This separation matters because Monzo operates in a regulated environment. The agent can run with room to develop and investigate, while its access to production infrastructure and data remains deliberately restricted.

### The agent became a reusable component for work beyond code generation
[08:33](https://www.youtube.com/watch?v=adfw3xJBx1o&t=513s)
Once Monzo connected Slack, GitHub, Linear, and its services to the agent platform, Agent Chip became a building block for new use cases. Engineers can delegate tasks to remote agents instead of supervising every step. The speakers say this increases the range and volume of work engineers can run and lets people spend more time on problems with ambiguity. The same component can support technology teams and business users without requiring a separate agent system for each task.

### Monzo uses Agent Chip for incidents, data quality, and service questions
[09:59](https://www.youtube.com/watch?v=adfw3xJBx1o&t=599s)
During an incident, Agent Chip joins the relevant Slack channel and investigates logs, traces, the affected service, and its code. Monzo wants it to provide high-signal support rather than flood the channel with hypotheses, and humans control how much help it provides. In ask channels, the agent can draft answers about backend services, products, or roadmaps. Agent Chip on Call investigates data quality failures and can proactively inspect backend changes for possible effects before the next batch processes data.

### Non-engineers can propose tested changes while engineers retain review responsibility
[12:07](https://www.youtube.com/watch?v=adfw3xJBx1o&t=727s)
Non-specialists can call Agent Chip from Slack when they have product feedback or an idea. The agent can produce a draft that has already been tested in staging, after which an engineer reviews and deploys the change. The speakers say this lowers the barrier for people outside specialist engineering roles to contribute to Monzo's systems. It does not remove Monzo's development standards or the engineer's role in checking and releasing the work.

## Notable quotes
- Felipe Felisola Caso: "We want to share the secret sauce of what makes Agent Chip a successful product and it's LLMs. But actually LLMs in this case stands for lots of little multipliers." (03:40)
- Felipe Felisola Caso: "The main idea here is that we want to separate this from the rest of our infrastructure in production." (06:46)
- Felipe Felisola Caso: "Now that Agent Chip is also a building block, the multiple parts of our bank can also leverage it to create other use cases." (08:57)
- Felipe Felisola Caso: "This doesn't change the rigor to which we develop software at Monzo, but this does allow a number of non-specialists to also start contributing more closely to the team." (12:50)
- Felipe Felisola Caso: "Our platform is provider agnostic by definition." (13:43)

## Tools & references mentioned
- Monzo
- Agent Chip
- MCP
- Slack
- GitHub
- Linear
- Anthropic

## Who should watch
- You are building an internal coding agent and need to decide which platform, access-control, and sandbox systems should exist before broad adoption.
- Your engineers spend time supervising agents or answering recurring service and architecture questions, and you want to delegate more work without removing review.
- You work in a regulated environment and need a practical pattern for giving agents useful internal context while restricting tools, permissions, and network access.

## Related talks

- [40x More Merged PRs With 2.7x the Engineers](https://aietalks.com/talks/40x-more-merged-prs-with-2-7x-the-engineers) (Walden Yan, Cognition, 14:08)
- [Containing Agent Chaos](https://aietalks.com/talks/containing-agent-chaos) (Solomon Hykes, Dagger, 23:48)
- [Agents for Everything Else](https://aietalks.com/talks/agents-for-everything-else) (swyx, AI Engineer, 14:10)
- [300 Agents, 60 People: How We Secured Our Own AI Agents](https://aietalks.com/talks/300-agents-60-people-how-we-secured-our-own-ai-agents) (Aiden Sochowski, Runlayer, 16:38)
- [How We Solved Agent Building](https://aietalks.com/talks/how-we-solved-agent-building) (Andrew Qu, Vercel, 17:34)
