- Monzo's engineers argued their coding agent worked quickly because the bank already had strong platform tooling, such as staging, deployment and alerting.
- They also said the agent runs in a sandbox with controlled access, and that it does not change the rigor of how Monzo develops software.
Two Monzo engineers, Felipe Felisola Caso and Francesco Galletta, argued that their bank's in-house coding agent took off mainly because of the platform Monzo already had. They made the case on the AI Engineer show. In their telling, the AI is one part of the story. The groundwork around it is the larger part.
What was said
The show notes say the agent began as an emoji you could add to a Slack message. They add that it now writes more than a tenth of the merged pull requests (finished code changes) across Monzo's tech teams. They also put its daily workload above 1,800 tasks.
The speakers said they were not there to brag. They wanted to explain why it worked. A speaker on the show said: "LLMs, in this case, stands for lots of little multipliers."
They listed what Monzo already had. That included alerting, monitoring, links to third-party services, a staging environment for easy testing, queuing, simple deployment and database tools. They said this platform let a simple Slack prompt turn into real work.
They then built a gateway using MCP, a standard way for agents to reach tools and data. It decides what the agents can see and handles who is allowed in. Agents run in a sandbox, kept apart from production systems. A speaker on the show explained why: "Again, we are a bank. We live in a regulated environment."
The agent itself then became a building block. It now joins incident channels, reads logs, traces and code, and aims to give humans a few useful signals rather than a stream of guesses.
Why it matters
Our reading is that this talk shifts where the buying question sits. Agent speed depends on things an agent can plug into: a safe place to test, a simple way to deploy and alerts that work. Monzo's engineers credited exactly those.
For leaders weighing an AI coding tool, the practical check is on your own base. Can an agent reach your systems safely? Can you test its work somewhere that is not production? The talk does not show what happens when those answers are no. Our reading is that this is the gap to examine before expecting similar results.
The other side
This is one company's account, told by the people who built the agent. The excerpts do not compare Monzo with firms that lack such a platform. They give no cost or timeline for building it. They do not report error rates or the quality of the agent's code.
The show notes say only that the agent authors over a tenth of merged pull requests, so we cannot say how much of Monzo's merged work still comes from people or other tools. And Monzo is a regulated bank, so some of its discipline may come from that setting rather than being a model others can copy.
The speakers did stress one limit. A speaker on the show said: "so this doesn't change the rigor to which we develop software at Monzo". The agent is meant to fit the existing standards, not replace them.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: A Bank Built Its Own Coding Agent. It Took Off — Felipe Felisola Caso & Francesco Galletta, Monzo (2026-10-20)
