Inside theSoftware Factory
A Monzo card being tapped on a card reader.

Photo: Monzo

fintech · UK · 1000+

Monzo logoInside Monzo's software factory

How Monzo's Agent Chip writes about 10% of merged PRs and runs 1,800+ tasks a day, through an in-house MCP gateway and sandboxed containers.

Monzo at a glance

The company

Sector"Digital bank" 6
EmployeesMore than 5,000 6
Customers15.2m personal and business customers 6
Revenue£1.7bn revenue (FY2026) 6

Monzo is regulated by the FCA and PRA. Customer count includes 14.2m personal and 905k business accounts 6.

The engineering organisation

Engineering size1,826 average headcount in Business and Technology (FY2026) 6
AI tooling sinceLate 2025, when the decision to build Agent Chip was made 1
Tool policyBuilt in-house explicitly to control guardrails as a regulated bank 1
Agents in the SDLCBackground coding agent; incident response; automated code review; non-engineer contribution 1
Reported adoption~10% of merged PRs authored by Agent Chip; 1,800+ tasks/day; 87% of tech teams using AI tooling weekly; 22% of code committed via AI tools 1,6
Published metricsPR share, daily task count, AI tooling adoption rate, code commit share 1,6

The 22% code commit figure comes from the FY2026 annual report and covers all AI coding tools, not just Agent Chip. The ~10% PR figure is specific to Agent Chip 1,6.

~10%of merged PRs authored by Agent Chip
1,800+tasks per day
87%of tech teams using AI tooling weekly
22%of code committed via AI tools
AI coding tools code commit share: 22%6 (Mar '26, current)
merged PRs authored by Agent Chip: ~10%1 (Aug '26, current)

People

Fabien Deshayes
Fabien Deshayes Engineering Manager
Oli Haley
Oli Haley Senior Backend Engineer

What they set out to solve

Regulation drove the build-vs-buy decision, and Monzo stated it directly. The post frames the central requirement as observability and control: "We're a bank, and observability and control is paramount. We need to know exactly what the agent can see, what it can do, and constrain it as appropriate" 1.

The decision to build was made in late 2025 "knowing that building gives us full control over the guardrails we need and the risks we need to manage" 1. This is the clearest causal statement in the corpus connecting compliance pressure to agent architecture.

A separate problem emerged once AI tooling started driving output: code review became the bottleneck. The response was not more AI review but isolated preview environments, called tenancies, for testing changes 2.

Maturity level

L4 · AI-nativeof 5
TraditionalAI-supportedAI-assistedAI-nativeAI-autonomous

Assigned by testing published facts against fixed criteria, and capped by what the sources state.

How Monzo's software factory works

Agent Chip is an omnipresent tool available across Slack, Linear, GitHub, and locally on a user's machine. It gives access to the vast majority of Monzo's repositories, documentation, and observability systems at a click or a one-sentence prompt 1.

The stated design goals were expanding the breadth and volume of tasks engineers can run, lowering the barrier for non-specialists to contribute, and augmenting the software delivery lifecycle by shifting routine work to agents while humans focus on higher-value problems 1.

Architecture

The high-level architecture has a series of inputs (Linear, Slack, GitHub) feeding into remote code agent services, which orchestrate an "agent bundle" deployed into a sandboxed environment isolated from the services that run the bank 1.

The bundle isolates an agent sufficiently to pick and choose what access it has on a per-task basis. The agent runs on an environment-specific Docker image depending on the job assigned and is given a single isolated proxy to speak with over HTTP. This allows secret isolation and de-risks agent access to the outside world 1.

Reliability is at the heart of the design. If a model provider suffers an outage or its performance degrades, Monzo can quickly swap in another. This prevents vendor lock-in and ensures tools are always available 1.

Fig. 1
Slack Linear GitHub Agent Chip orchestrator Sandbox Docker image, isolated proxy, per-task tools MCP gateway knowledge sources Pull request 👤 Isolated from production; secrets never enter the agent environment [1] A human always reviews and merges every Agent Chip pull request [1]
How Agent Chip runs a task through sandboxed containers, as described in the Building Agent Chip post [1]. Original diagram, not a reproduction.

MCP gateway

An MCP gateway provides centralized access to knowledge sources that inform models about patterns, practices, and learnings, enriching agent context with relevant information for better outputs 1.

Where it's been most useful

Three use cases stand out 1:

Incidents. When something breaks, the cost of a slow first response is high. Agent Chip pulls together context across systems, surfaces likely causes, and can draft a fix while humans coordinate the response. Monzo has seen Agent Chip's Incident Mode catching errors from responding engineers that could have extended resolution time.

Non-engineers shipping small fixes. Some of the most valuable changes to products aren't deeply technical: copy tweaks, small UX inconsistencies, edge cases caught by customer operations. With Agent Chip, people closest to the problem can describe a fix in plain English and get a pull request out of it. An engineer still reviews and merges, but the bottleneck moves from "find an engineer with capacity" to "find a reviewer."

Routine engineering work. Library upgrades, repetitive refactors, the kind of work that's important but rarely the most exciting thing on someone's plate. Agent Chip handles this work, and the time saved compounds.

Federated adoption

To enable fast iteration, Monzo designed for federated adoption by separating Agent Chip "the platform" from Agent Chip "the product." A trigger service consumes events from internal systems (Slack messages, Linear tasks, webhooks) and maps them onto workflows. Workflows define throughput type (batch job or ASAP), response structure, and response handling. This lets teams across the business quickly set up Agent Chip integrations, and many integrations have been created by Agent Chip itself 1.

Operating Monzo's software factory: teams, roles, and skills

Enablement became a platform responsibility, not individual choice. This was explicitly attributed to regulation 2. Compare Zalando and Block, where tool choice is deliberately left to teams.

Adoption was staged as an experiment. Monzo started broad and unstructured, then switched to small cohorts with defined use cases and explicit success metrics before wider rollout, treated as platform governance rather than individual enthusiasm. A cross-department AI Engineering Champions network does peer-driven tool testing 2.

Spend is capped and reviewed: roughly $1,000 per engineer per year, revisited quarterly with CTO and finance alignment 2. This is the opposite of Ramp's deliberately uncapped approach.

Implementation journey

  1. Late 2025
    Decision made to build Agent Chip in-house to control guardrails as a bank 1.
  2. Late 2025 + ~3 weeks
    Agent Chip MVP shipped 1.
  3. October 2025
    Podcast interview describes AI tooling budget, small cohort approach, and code review as the new bottleneck 2.
  4. May 2026
    Reimagining monzo.com post describes design-to-code via design-tool MCP 3.
  5. August 2026
    Building Agent Chip post published: ~10% of merged PRs, 1,800+ tasks/day 1.
  6. FY2026
    Annual report shows 87% of tech teams using AI tooling weekly, 22% of code committed via AI tools 6.

Measuring Monzo's software factory

The engineering blog reports ~10% of merged PRs authored by Agent Chip and 1,800+ tasks running daily 1. These are self-reported with no disclosed definition of what counts as agent-authored or as a task.

The annual report provides broader figures: 87% of technology teams using AI tooling every week, and 22% of all code committed by AI coding tools 6. The 22% figure covers all AI coding tools, not just Agent Chip.

The podcast provides additional context: PRs per engineer up 10–20%, average PR size up similarly, ~20% of new code AI-generated 2. These figures come from a developer-productivity vendor's platform.

On the customer operations side, AI self-service handles 62% of inbound customer chats and resolves 55% independently, with a 4.4/5 customer satisfaction score 6.

Lessons learned

The answer to the review bottleneck was environments, not more AI. Ramp, Uber, Block, and Shopify all responded to rising PR volume by adding automated review. Monzo added tenancies, isolated preview environments for testing changes. Same diagnosis, structurally different fix 2.

Scoping is per-task, not blanket. Each agent task runs in a sandboxed container with a restricted, per-task subset of tools. A human always reviews and merges 1.

The operations agent's design appears to have been reused as a template for Agent Chip's guardrails: input/output guardrails, triage logic, knowledge-base grounding, and human escalation when confidence is low 4.

What's next for Monzo's software factory

The post describes a six-month roadmap 1:

  • Making Agent Chip even better at supporting engineers: accessing the entire knowledge base and building, running, testing, and deploying across the entire codebase.
  • Making it even easier for non-engineers across the company to use Agent Chip for things that previously required an engineer.
  • Enhancing the review, evaluation, and safety infrastructure that gives confidence in Agent Chip's capabilities, including PR review.

Patterns in this factory

Gaps in the record (5)
  • No definition of what counts as "agent-authored" or as a "task" for the reported figures.
  • No stated cost or spend figures for Agent Chip beyond the ~$1,000/engineer/year AI tooling budget mentioned in a podcast.
  • The 22% code commit figure is from the annual report and covers all AI coding tools, not just Agent Chip; the two figures are not directly comparable.
  • The podcast figures (10–20% PR increase, ~20% of new code AI-generated) come from a developer-productivity vendor's platform.
  • No disclosed definition of how the risk-tier approval rules work in detail.
Sources (6)
1
Building Agent Chip: Monzo's in-house agentic tool
Fabien Deshayes, Oli Haley · Engineering Manager; Senior Backend Engineer · 2026-08-13
2
How Monzo runs data-driven AI experimentation
Fabien Deshayes · platform engineering leader · 2025-10-31
3
Reimagining monzo.com
Lyudmyla Maksymchuk, Weijie Fu · Lead Product Designer; Senior Product Designer · 2026-05-27
5
Monzo
Monzo · corporate disclosure
6
FY26 MBHG Consolidated Annual Report
Monzo Bank Holding Group Limited · corporate disclosure · 2026-05