Monzo built its own coding agent and 10% of PRs followed

AI Engineer captured how Monzo engineers Felipe Felisola Caso and Francesco Galletta built Agent Chip and watched it take over work. The agent now authors over 10% of all merged PRs across tech and runs more than 1,800 tasks per day, the pair said.

The bank runs more than 3,000 backend services and around 13,000 data warehouse models, plus dozens of ML predictors for credit and financial crime, they said. In January 2026 a Slack channel went viral after the team added an Agent Chip emoji, Felisola Caso said. React with the emoji and any message becomes a prompt, and the team then made it an always available teammate in Slack, GitHub and Linear.

What made it stick was not the model, the pair argued. Monzo describes itself as a bank that lives on your phone, according to its Google Play listing, and that mobile product runs on a platform the team could reuse for agents. Felisola Caso said the platform already had alerting, monitoring, staging, queuing, deployment and database abstractions, so the emoji-to-prompt trick had somewhere to run. One line summed the thesis: LLMs mean lots of little multipliers.

The architecture is deliberately boring.

First came an MCP gateway, Galletta said, to give local agents deterministic access to internal services and knowledge sources with auth handled in one place. Then came orchestration for remote agents. Monzo runs agents in an isolated namespace, the pair said, with multiple sandbox pods. Each agent image defines its CLIs and permissions, a proxy controls egress to code and repos, and the agent itself runs in what they called a box inside a box where nondeterminism is contained.

That separation matters because Monzo is regulated, they said. Production stays apart from the sandbox while the agent can still read logs, traces and code. On top of those blocks the team wired Slack, GitHub, Linear and the more than 3,000 services, the pair said. The little blue-green agent box can be swapped for any provider, so the system is model agnostic, they said, noting they could have joked they built Claude Code six months before Anthropic shipped it.

New use cases now compose quickly, they said. In incident mode Agent Chip joins the channel, pulls logs and traces and drafts a high-signal investigation for a human. In ask channels it answers squad-to-squad questions from knowledge sources and drafts replies. A newer Agent Chip on Call triages data quality alerts, and a proactive variant scans backend changes before the next batch to catch data breakage early.

The same entry point lowered the bar for non-specialists, they said. Product and ops staff can ask Agent Chip in Slack for a draft, get it tested in staging, then loop in an engineer for review. The pair said coverage still depends on controlled egress and human review, and they gave no timeline for expanding from backend and app drafts to data tests and predictors.