Monzo’s new AI agent is now writing 10% of the bank’s code

Monzo’s in-house coding agent, Agent Chip, now authors around 10% of all merged pull requests at the bank and runs more than 1,800 tasks every day.

The digital bank set out the numbers in a blog post by Fabien Deshayes, engineering manager at Monzo, and Oli Haley, senior backend engineer at Monzo.

Nearly all of Monzo’s engineers already use at least one coding agent as part of their daily work, according to the post, after the bank trialled a range of third-party tools over the past 18 months.

Monzo decided in late 2025 to build its own agent rather than rely on an off-the-shelf product, and it reached a minimum viable product in around three weeks.

Why the bank built its own platform

Monzo said owning its harness gave it full control over what the agent can see and do, which it treated as a requirement rather than a preference given it is a regulated bank.

The bank also said the setup lets it swap in a different model if a provider suffers an outage or its performance degrades, which avoids vendor lock-in and keeps the tooling available.

Two pieces of foundational infrastructure sit underneath the product. The first is an MCP gateway that gives the agent centralised access to Monzo’s internal knowledge sources, patterns and practices.

The second is remote agentic infrastructure that spawns a sandboxed container for each job, well away from the services that actually run the bank.

Each agent runs inside an isolated bundle with a single proxy for outbound HTTP requests, which lets Monzo choose the agent’s access on a per-task basis and keep secrets away from it.

What Agent Chip actually does

Agent Chip is available across Slack, Linear, GitHub and on an engineer’s own machine, with access to the majority of Monzo’s repositories, documentation and observability systems.

The current setup executes single, one-shot jobs, including:

• Alert investigation: Automatic investigations of paging alerts in production
• Prompt via Slack: Implementing pull requests, bug fixes and reactive work from Slack context
• Team responder: Answering inter-team questions by searching the codebase and knowledge base
• Code reviews: Automated code reviews in GitHub
• Incident responder: A live agent that monitors an incident and helps with root cause analysis

Monzo said its incident mode has caught errors from responding engineers on several occasions, mistakes that could have complicated resolution or extended the time to fix.

The bank also pointed to non-engineers shipping small fixes as one of the most valuable use cases so far. Copy tweaks, minor interface inconsistencies and edge cases spotted by customer operations can now start as a plain-English description and end as a pull request.

An engineer still reviews and merges that work, so the constraint moves from finding an engineer with spare capacity to finding a reviewer.

Routine engineering work makes up the third bucket, covering library upgrades and repetitive refactors that Monzo described as important but rarely the most interesting job on an engineer’s plate.

Jobs run through an entry point called remote-code-agent-trigger, which consumes events from Slack messages, Linear tasks and webhooks, then maps them onto workflows.

Those workflows set the throughput type, the structure of the agent’s response and how Monzo handles it, which lets the bank attribute jobs to teams and monitor cost at a product level.

Teams across the business have set up their own Agent Chip integrations off the back of that, and Monzo said the agent has built many of those integrations itself.

The bank plans to give Agent Chip access to its entire knowledge base next, along with the ability to build, run, test and deploy across the full codebase.

Monzo is also expanding the review, evaluation and safety infrastructure around the agent, including pull request review, and is hiring into the team behind it.

Now read: Bill Gates wants a tax on robots and AI tokens