Monzo’s Felipe Felisola Caso and Francesco Galletta explain their in-house coding agent in this 15-minute talk on AI Engineer.

From a Slack reaction to delegated work

  • React to a Slack message with an emoji, and Agent Chip takes that message as its prompt.
  • Engineers also use mentions, code reviews and Linear tickets to hand it work.
  • The speakers report that it authors over 10% of merged pull requests across Monzo’s technology teams and runs more than 1,800 tasks daily.
  • Their setting includes over 3,000 back-end services, around 13,000 data warehouse models and dozens of machine-learning predictors.

The existing platform did much of the work

  • Their expansion of “LLMs” is “lots of little multipliers”: monitoring, alerting, integrations, staging, queues, deployment and database abstractions.
  • An MCP gateway, a controlled connection between agents and tools, exposes selected internal services and knowledge sources.
  • The gateway manages authentication and authorization for both local coding assistants and remote agents.

Put permissions outside the model

  • An orchestration system runs agents in sandboxes, isolated environments separated from the bank’s production infrastructure.
  • Agent images define available tools and command-line programs; the gateway supplies controlled access.
  • A proxy restricts outbound network access to deliberately permitted destinations, including repositories.
  • The model operates inside those boundaries, while surrounding software controls how it connects to the bank.

Reuse the same service beyond coding

  • Incidents: inspect logs, traces, affected services and code, while humans control how much the agent participates.
  • Team questions: draft answers from internal knowledge, ask for missing context and bring in a human when needed.
  • Data quality: investigate failed jobs reactively, or inspect back-end changes for possible downstream trouble before the next batch runs.
  • Non-engineers: turn product feedback into a draft tested in staging, then involve engineers for review and deployment.

Why keep building it in-house?

  • The controlled tool access and execution environment become reusable infrastructure for additional workflows.
  • The speakers describe the model component as replaceable, letting Monzo choose providers by task.
  • Broader end-to-end testing across app, back-end and data changes is a direction for expansion.
  • Caveat: the adoption figures are Monzo’s own report. The talk gives no controlled productivity comparison, defect-rate analysis or security audit.

“LLMs in this case stands for lots of little multipliers.” — 4:00