A Bank Built Its Own Coding Agent. It Took Off — Felipe Felisola Caso & Francesco Galletta, Monzo

Highlights

• Agentship currently authors over 10% of all merged pull requests across engineering, data, and machine learning at Monzo.

• The internal coding agent handles more than 1,800 tasks per day within the bank's technical ecosystem.

• Engineers maintain over 3,000 backend services, 13,000 data warehouse models, and dozens of machine learning predictors.

• An MCP gateway provides deterministic access to internal tools and knowledge sources while managing authorization.

• Remote agents run inside isolated pods with specific environment configurations, CLIs, and restricted egress proxies.

Timeline

1. 00:00:00 - 03:28:00

   Scaling Infrastructure and Introducing Agent Chip

   Operating as a digital bank requires managing complex use cases across financial crime, credit decisions, and app backends. To handle this scale, internal teams adopted an agent harness accessible directly through Slack interactions. This tool evolved from a simple prompt receiver into an always-available coworker that merges pull requests and executes thousands of daily workflows.

   • Monzo maintains over 3,000 backend services, 13,000 data warehouse models, and multiple machine learning predictors.
   • An internal Slack tool called Agent Chip triggers tasks simply through emoji reactions.
   • The agent functions as an active teammate capable of answering questions, performing code reviews, and executing linear tasks.
2. 03:28:00 - 08:36:00

   Platform Capabilities and Architecture

   Building successful internal agents relies heavily on existing platform infrastructure rather than raw model capabilities alone. A dedicated MCP gateway manages authentication and exposes selected services safely to the models. Furthermore, an orchestration system separates non-deterministic agent behavior from production environments by running tasks inside isolated namespace pods with controlled proxies.

   • A strong underlying platform with alerting, staging environments, and database abstractions enabled rapid agent development.
   • An MCP gateway connects local and remote environments to internal knowledge sources and services with deterministic behavior.
   • Remote agents operate inside isolated Kubernetes namespaces within sandboxes equipped with strict tool permissions.
3. 08:36:00 - 15:21:00

   Business Outcomes and Operational Use Cases

   Composability allows different parts of the bank to leverage Agent Chip as a core building block for new use cases. The system handles reactive tasks such as triaging data quality alerts and active incident investigations, alongside proactive monitoring of backend changes. Additionally, non-technical business users leverage the interface to draft queries and retrieve roadmap context with lower barriers to entry.

   • Remote agents automate routine processes, enabling engineers to focus on high-value problems involving ambiguity.
   • Agent Chip automatically joins incident channels to analyze logs and traces during active system failures.
   • Non-specialists utilize the platform to draft and test changes in staging environments before involving engineers for final deployment.