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

Um, we... Well, that was Um, we... Well, that was loud. Um, we are Felipe loud. Um, we are Felipe loud. Um, we are Felipe and Francesco. Um, we're here to and Francesco. Um, we're here to and Francesco. Um, we're here to talk to talk to talk to you about the cumulative you about the cumulative you about the cumulative value of building value of building value of building Monzo Sync House agents Monzo Sync House agents Monzo Sync House agents with LLM. Um, so what is with LLM. Um, so what is with LLM. Um, so what is Monzo? We are a bank. A Monzo? We are a bank. A Monzo? We are a bank. A digital bank that digital bank that digital bank that strives to make strives to make strives to make money accessible to money accessible to money accessible to everyone. Um, to everyone. Um, to everyone. Um, to give you an idea of ​​the scale of the give you an idea of ​​the scale of the agent coding problem: agent coding problem: we have teams of we have teams of we have teams of engineers engineers engineers supporting over 3,000 supporting over 3,000 supporting over 3,000 backend services, backend services, backend services, about 13,000 about 13,000 about 13,000 data warehouse models, and data warehouse models, and data warehouse models, and dozens and dozens of dozens and dozens of ML prediction models. So ML prediction models. So we do everything we do everything we do everything from credit decision-making from credit decision-making from credit decision-making and and and financial crime financial crime financial crime to implementing very to implementing very to implementing very complex tasks for complex tasks for complex tasks for the application, backend, the application, backend, the application, backend, and data. Um, so I'm and data. Um, so I'm and data. Um, so I'm sure even sure even sure even today, outside of today, outside of today, outside of this conference, you've this conference, you've this conference, you've heard a lot of heard a lot of heard a lot of success stories of companies success stories of companies success stories of companies implementing AI implementing AI agents for agents for agents for programming. Um, programming. Um, programming. Um, actually, a couple of weeks actually, a couple of weeks actually, a couple of weeks ago there was a conference here in New York, ago there was a conference here in New York, ago there was a conference here in New York, and and and this article came out this article came out this article came out about the state of the about the state of the about the state of the tech tech tech industry in 2026.

industry in 2026. industry in 2026. Um, and there was this, Um, and there was this, Um, and there was this, hopefully hopefully hopefully humorous, quote from X humorous, quote from X humorous, quote from X about how if you don't about how if you don't about how if you don't build your own build your own build your own AI coding system, AI coding system, AI coding system, can you can you can you even be considered a even be considered a even be considered a serious serious serious tech company? So there's something to tech company? So there's something to tech company? So there's something to think about, think about, think about, right? Um, then right? Um, then right? Um, then inside Monzo, on January inside Monzo, on January inside Monzo, on January 26th, this Slack channel 26th, this Slack channel 26th, this Slack channel went viral. Agent Chip. went viral. Agent Chip. went viral. Agent Chip. And I hope I can And I hope I can ... Oh no, what happened? You ... Oh no, what happened? You ... Oh no, what happened? You can't can't can't do this. Everything is fine. do this. Everything is fine. do this. Everything is fine. So hopefully I So hopefully I So hopefully I can draw can draw can draw your attention to this your attention to this your attention to this little emoji little emoji little emoji here. Very cute. We here. Very cute. We here. Very cute. We adore him. Um, this adore him. Um, this adore him. Um, this went very, very went very, very went very, very viral within the viral within the viral within the company for a very company for a very company for a very good reason. You good reason. You good reason. You can give him can give him any prompt by simply any prompt by simply any prompt by simply responding to a responding to a responding to a message with this message with this message with this little emoji, and little emoji, and little emoji, and he'll pick up on it as, he'll pick up on it as, he'll pick up on it as, you know, a request. Um, so you know, a request. Um, so you know, a request. Um, so it evolved, we it evolved, we it evolved, we doubled down on it, and doubled down on it, and doubled down on it, and it became a constantly it became a constantly it became a constantly available colleague in available colleague in available colleague in our Slack and throughout the our Slack and throughout the our Slack and throughout the work environment. work environment. work environment. Um, to give Um, to give Um, to give a few examples: he a few examples: he a few examples: he really feels really feels really feels like a colleague. Um, he's like a colleague. Um, he's like a colleague. Um, he's just nearby. You just nearby. You just nearby. You can always can always can always tag Agent Chip to tag Agent Chip to tag Agent Chip to do something for do something for do something for you: answer you: answer you: answer a question, learn a question, learn a question, learn something about architecture, something about architecture, something about architecture, or just or just or just chat, although I chat, although I chat, although I don't think he don't think he don't think he likes it very much. Um, likes it very much. Um, likes it very much. Um, he can also do he can also do he can also do code reviews. So, as a code reviews. So, as a code reviews. So, as a very experienced very experienced very experienced engineer, he engineer, he engineer, he will simply tell you that everything will simply tell you that everything will simply tell you that everything looks good and you looks good and you looks good and you calmly move on.

calmly move on. You can also You can also assign it assign it assign it linear tasks. So linear tasks. So linear tasks. So you can give him you can give him you can give him tickets and he'll just tickets and he'll just tickets and he'll just get to get to get to work. Um, so that's all work. Um, so that's all work. Um, so that's all well and good. well and good. Fast forward to Fast forward to today: today: today: Agent Ship generates over 10 Agent Ship generates over 10 % of all merged PR in the % of all merged PR in the % of all merged PR in the technology technology technology department. This department. This department. This applies to data, applies to data, applies to data, machine learning, machine learning, machine learning, backend engineering, backend engineering, backend engineering, applications, etc. It applications, etc. It applications, etc. It also completes over also completes over also completes over 1,800 tasks per day, which is quite 1,800 tasks per day, which is quite 1,800 tasks per day, which is quite a lot. And it is also a lot. And it is also a lot. And it is also starting to be actively starting to be actively starting to be actively adopted by adopted by adopted by business users business users . That is, it is not only . That is, it is not only . That is, it is not only programming and programming and programming and technology, it also technology, it also technology, it also tries tries tries to answer to answer to answer business questions, business questions, business questions, write queries, or write queries, or write queries, or independently provide independently provide independently provide everything a everything a everything a business business user might need. So, user might need. So, user might need. So, again, you've heard again, you've heard again, you've heard a lot of a lot of a lot of success stories about companies success stories about companies success stories about companies implementing AI implementing AI agents for agents for agents for coding, but what of coding, but what of coding, but what of it? Why are we...what are we it? Why are we...what are we it? Why are we...what are we trying trying trying to tell today? to tell today? to tell today? Do we want to Do we want to Do we want to brag about having brag about having brag about having built Claude Tag 6 built Claude Tag 6 built Claude Tag 6 months before months before months before Anthropic released it?

Anthropic released it? Anthropic released it? Perhaps. But that is not the Perhaps. But that is not the Perhaps. But that is not the focus of focus of focus of today's today's today's conversation. We want conversation. We want conversation. We want to share the secret to share the secret to share the secret ingredient of what ingredient of what ingredient of what makes Agent Ship a makes Agent Ship a makes Agent Ship a successful product— successful product— it's LLM. But in fact, LLM it's LLM. But in fact, LLM it's LLM. But in fact, LLM in this case in this case in this case means " means " lots of little multipliers". We lots of little multipliers". We believe that what believe that what believe that what allowed us allowed us allowed us to build Agent Ship so to build Agent Ship so to build Agent Ship so quickly and so quickly and so quickly and so efficiently is the very efficiently is the very efficiently is the very powerful platform at powerful platform at powerful platform at Monzo. So when we Monzo. So when we Monzo. So when we say it went say it went say it went viral and you could viral and you could viral and you could write a prompt, write a prompt, write a prompt, respond, and respond, and respond, and make a profit. make a profit. make a profit. In fact, it is the In fact, it is the In fact, it is the result of many result of many result of many opportunities that we opportunities that we opportunities that we already had. These are already had. These are already had. These are notifications, notifications, notifications, monitoring, monitoring, monitoring, integrations with integrations with integrations with third-party services third-party services like the ones like the ones like the ones Francesco mentioned, a Francesco mentioned, a Francesco mentioned, a staging environment staging environment staging environment for easy for easy for easy testing, queues, testing, queues, testing, queues, easy deployment, easy deployment, easy deployment, database abstractions, database abstractions, database abstractions, and many other things and many other things I could I could I could mention. So, mention. So, mention. So, first of all, it's what first of all, it's what first of all, it's what we have in our we have in our we have in our platform that has allowed platform that has allowed platform that has allowed Agent Ship to become successful.

Agent Ship to become successful. But it doesn't But it doesn't end there. Even end there. Even end there. Even before Agent before Agent before Agent Ship emerged as a remote Ship emerged as a remote Ship emerged as a remote agent, we started agent, we started agent, we started investing in investing in investing in supporting AI agents supporting AI agents supporting AI agents for coding even for coding even for coding even in on-premises in on-premises in on-premises environments. So environments. So environments. So one of the first things one of the first things we did was we did was we did was develop an MCP gateway develop an MCP gateway develop an MCP gateway that provided easy that provided easy that provided easy access to a set of access to a set of access to a set of knowledge sources, knowledge sources, knowledge sources, internal internal internal tools, and tools, and tools, and some of the services some of the services some of the services that that that Francesco mentioned. This is what Francesco mentioned. This is what the architecture looks like. We have the architecture looks like. We have things we want things we want things we want to connect as to connect as to connect as internal services and internal services and internal services and knowledge sources. We knowledge sources. We knowledge sources. We put this behind an MCP put this behind an MCP server that we server that we server that we created to created to created to get get get deterministic deterministic deterministic behavior, choosing behavior, choosing behavior, choosing what exactly we want to what exactly we want to what exactly we want to provide to the agents. provide to the agents. provide to the agents. Additionally, we manage Additionally, we manage Additionally, we manage authorization and authorization and authorization and authentication for authentication for authentication for many different many different many different agents. Again, agents. Again, agents. Again, this started in a world of this started in a world of this started in a world of local local local code writing agents code writing agents code writing agents looking to looking to looking to connect. By now, connect. By now, connect. By now, everyone has probably everyone has probably everyone has probably heard in numerous heard in numerous heard in numerous sessions that the more sessions that the more sessions that the more context and data we context and data we context and data we can safely can safely can safely provide to our agents, provide to our agents, provide to our agents, the better. After the better. After the better. After that, we created an that, we created an agent orchestration system that actually agent orchestration system that actually runs Agent runs Agent runs Agent Ship behind the scenes in a Ship behind the scenes in a Ship behind the scenes in a sandbox with very sandbox with very sandbox with very strict strict strict tools and tools and tools and permissions. So, if permissions. So, if permissions. So, if we go back to we go back to we go back to our diagram, the entire our diagram, the entire our diagram, the entire MCP MCP server architecture actually server architecture actually server architecture actually boils down to this boils down to this boils down to this block, where we block, where we block, where we plug in the rest of the

plug in the rest of the plug in the rest of the Agent Ship architecture. Agent Ship architecture. So, we have So, we have remote agents remote agents remote agents running in a sandbox running in a sandbox running in a sandbox in an isolated in an isolated in an isolated namespace, and the namespace, and the namespace, and the orchestration system orchestration system orchestration system runs, reads, runs, reads, runs, reads, and retrieves this and retrieves this and retrieves this information from there. The information from there. The main idea here—and main idea here—and this this this is emphasized again by is emphasized again by is emphasized again by many many many providers and providers and providers and labs— labs— is that we is that we is that we want to separate want to separate want to separate this from the rest of our this from the rest of our this from the rest of our production production production infrastructure. infrastructure. Again, we are a bank , we operate in a , we operate in a , we operate in a regulated regulated regulated environment, so environment, so environment, so we want to make sure we want to make sure we want to make sure that that that the non-determinism of the non-determinism of the non-determinism of agents agents agents is controlled. But is controlled. But is controlled. But at the same time, we want to at the same time, we want to at the same time, we want to have an environment where have an environment where have an environment where we can allow we can allow we can allow them to work them to work them to work and develop freely. And and develop freely. And and develop freely. And again, if we again, if we again, if we expand on this to expand on this to expand on this to give you a little more give you a little more give you a little more understanding of what's understanding of what's understanding of what's going on "under going on "under going on "under the hood." So, in the hood." So, in the hood." So, in this isolated this isolated this isolated namespace, namespace, namespace, multiple multiple multiple sandboxed pods are created, sandboxed pods are created, sandboxed pods are created, on which on which on which agents run. The agent image agents run. The agent image is where we is where we is where we define the exact define the exact define the exact tools that tools that tools that agents will have agents will have agents will have access to. These are things access to. These are things access to. These are things like the environment, like the environment, command line interfaces (CLIs), command line interfaces (CLIs), authentication, and authentication, and authentication, and authorization, which again we authorization, which again we get from the get from the gateway itself. A proxy is what gateway itself. A proxy is what gateway itself. A proxy is what allows us allows us allows us to connect to the to connect to the to connect to the outside world.

outside world. outside world. This is how, for example, This is how, for example, This is how, for example, Agent Chip can Agent Chip can Agent Chip can connect to connect to connect to our code, to our our code, to our our code, to our repositories, and repositories, and repositories, and work with them. But, work with them. But, work with them. But, of course, we of course, we of course, we only provide the outgoing only provide the outgoing only provide the outgoing traffic that we ourselves traffic that we ourselves traffic that we ourselves decide to allow. decide to allow. So, this is how we So, this is how we ensure that everything is ensure that everything is ensure that everything is very controlled, very controlled, very controlled, consciously granting consciously granting consciously granting access only to what access only to what access only to what we want to reveal we want to reveal . And finally, the little . And finally, the little . And finally, the little blue-green block blue-green block blue-green block there—that's where the there—that's where the there—that's where the agent himself works. agent himself works. agent himself works. So this block So this block So this block within a block within a block within a block within a block—this is within a block—this is within a block—this is where where the indeterminacy happens and the indeterminacy happens and where the agent can do whatever it where the agent can do whatever it where the agent can do whatever it wants. So, this is what the wants. So, this is what the wants. So, this is what the whole whole whole scheme looks like when we scheme looks like when we scheme looks like when we give our MCP give our MCP server access to server access to server access to everything that is important everything that is important everything that is important for the agent; we can for the agent; we can for the agent; we can securely give it securely give it securely give it access to our data access to our data , and the , and the , and the agent orchestration system agent orchestration system agent orchestration system takes care of the rest. takes care of the rest. Having these building Having these building blocks makes any blocks makes any blocks makes any subsequent use cases subsequent use cases much easier to create much easier to create . As . As . As Francesco mentioned, we can Francesco mentioned, we can Francesco mentioned, we can connect Slack, then connect Slack, then connect Slack, then GitHub, then Linear, and in a GitHub, then Linear, and in a GitHub, then Linear, and in a different color I different color I different color I marked over 3000 marked over 3000 marked over 3000 services that we services that we services that we can connect as an can connect as an can connect as an addition to this addition to this addition to this block. Just block. Just block. Just as we had as we had as we had several elements in several elements in several elements in our platform to our platform to our platform to build Agent Chip, build Agent Chip, build Agent Chip, now that Agent Chip itself now that Agent Chip itself now that Agent Chip itself has become a building

has become a building has become a building block, block, block, our bank's divisions can our bank's divisions can our bank's divisions can use it use it use it to develop other to develop other to develop other solutions. We see that solutions. We see that solutions. We see that these building blocks these building blocks these building blocks allow us to not only allow us to not only allow us to not only add internal add internal add internal product features, but also product features, but also product features, but also achieve specific achieve specific achieve specific business outcomes. business outcomes. We have seen We have seen an increase in the number an increase in the number an increase in the number and scope of tasks and scope of tasks and scope of tasks that engineers are completing that engineers are completing that engineers are completing because they because they because they no longer need to no longer need to no longer need to constantly monitor constantly monitor constantly monitor their agents. They their agents. They their agents. They can can can use use use their their their programming assistant for programming assistant for programming assistant for specific tasks, specific tasks, specific tasks, but also delegate but also delegate but also delegate them to remote them to remote them to remote agents. We can agents. We can automate automate routine processes much better. This routine processes much better. This routine processes much better. This allows people allows people allows people to focus on to focus on to focus on more important more important more important tasks where there is tasks where there is tasks where there is especially a lot of especially a lot of especially a lot of uncertainty. Here are uncertainty. Here are uncertainty. Here are some examples of what some examples of what some examples of what we can we can we can do. This is an Agent Chip that do. This is an Agent Chip that do. This is an Agent Chip that operates in operates in operates in incident mode and incident mode and incident mode and uses the MCP to uses the MCP to uses the MCP to obtain current obtain current obtain current data. As soon as we data. As soon as we data. As soon as we detect and detect and detect and launch a new launch a new launch a new incident, Agent Chip incident, Agent Chip incident, Agent Chip joins joins joins the channel and starts the channel and starts the channel and starts an investigation: an investigation: an investigation: analyzes logs, analyzes logs, analyzes logs, traces, traces, traces, connects to a connects to a connects to a specific service specific service specific service and code to and code to and code to understand what is understand what is understand what is happening and happening and happening and provide people with accurate provide people with accurate provide people with accurate information. We don't information. We don't information. We don't want the agent to be want the agent to be constantly constantly saying things or saying things or saying things or testing hypotheses in testing hypotheses in testing hypotheses in real time during an incident. We real time during an incident. We real time during an incident. We strive for high strive for high strive for high accuracy, so we accuracy, so we accuracy, so we give people give people give people the ability the ability the ability to control the level of to control the level of to control the level of support from Agent support from Agent support from Agent Chip. We can

Chip. We can Chip. We can continue to continue to continue to use it use it use it in Slack not only through in Slack not only through in Slack not only through mentions, but also in so- mentions, but also in so- mentions, but also in so- called "ask channels." called "ask channels." called "ask channels." Essentially, this is the interaction Essentially, this is the interaction Essentially, this is the interaction between teams between teams between teams within Mondzo. within Mondzo. The team needs The team needs context around the backend context around the backend service, service, service, product development, and product development, and product development, and roadmap, and roadmap, and roadmap, and because Agent Chip can because Agent Chip can because Agent Chip can connect to connect to connect to our knowledge bases, it is our knowledge bases, it is our knowledge bases, it is able to draft a able to draft a able to draft a response response response that the human can that the human can that the human can use as use as use as reference or support. reference or support. Perhaps this Perhaps this answer will be answer will be answer will be enough, and Agent Chip enough, and Agent Chip enough, and Agent Chip will resolve will resolve will resolve this ticket on its own. Sometimes it this ticket on its own. Sometimes it this ticket on its own. Sometimes it will ask for more will ask for more will ask for more context, and the person context, and the person context, and the person overseeing the process overseeing the process overseeing the process can be tagged to step can be tagged to step can be tagged to step in in in to help. to help. to help. We've We've We've been implementing something recently that's been implementing something recently that's been implementing something recently that's been very been very been very helpful: sorting helpful: sorting data quality alerts and checking data quality alerts and checking backend changes that backend changes that backend changes that might cause might cause might cause problems. We problems. We problems. We call it Agent Chip on call it Agent Chip on call it Agent Chip on Call: it tracks Call: it tracks Call: it tracks failures and automatically failures and automatically failures and automatically initiates initiates initiates an investigation. This is an investigation. This is an investigation. This is its reactive its reactive its reactive version. What if version. What if version. What if errors never errors never errors never occurred? Based on occurred? Based on occurred? Based on this, we created a this, we created a this, we created a proactive version proactive version proactive version that monitors backend that monitors backend changes and identifies changes and identifies changes and identifies potential impacts potential impacts potential impacts to our data before to our data before to our data before it is it is it is processed during the processed during the processed during the next next next batch job.

batch job. These are just examples of how our technical how our technical how our technical team is using team is using team is using this, but even better, we this, but even better, we this, but even better, we are seeing a are seeing a are seeing a lowering of the barrier for lowering of the barrier for lowering of the barrier for non-technical people in the non-technical people in the non-technical people in the company who company who company who can now contribute can now contribute can now contribute to our systems. With to our systems. With to our systems. With quick access to Agent quick access to Agent quick access to Agent Chip in Slack, Chip in Slack, Chip in Slack, non-specialists can non-specialists can call Agent Chip based on feedback, ideas, or discussions, call Agent Chip based on feedback, ideas, or discussions, get a working get a working get a working draft already draft already draft already tested in a tested in a tested in a staging environment, and staging environment, and staging environment, and only then engage only then engage only then engage an engineer to an engineer to an engineer to verify and verify and verify and properly properly properly deploy changes. This deploy changes. This deploy changes. This doesn't change the rigor doesn't change the rigor doesn't change the rigor with which we develop with which we develop software at Monzo, software at Monzo, but it allows but it allows but it allows many many many non-specialists non-specialists to work more closely with the to work more closely with the team. You might team. You might team. You might think: the market think: the market think: the market is developing very is developing very is developing very quickly. We hear quickly. We hear quickly. We hear a lot of interesting news a lot of interesting news a lot of interesting news about developments, new about developments, new about developments, new products and features products and features products and features from various companies from various companies from various companies here at the conference and here at the conference and here at the conference and beyond. Why do we beyond. Why do we beyond. Why do we believe Agent Chip believe Agent Chip believe Agent Chip remains a valuable remains a valuable remains a valuable tool in tool in tool in 2026 and beyond 2026 and beyond ? As I said before and I want ? As I said before and I want ? As I said before and I want to repeat, because to repeat, because to repeat, because Agent Chip is composable, Agent Chip is composable, Agent Chip is composable, it has become a new it has become a new it has become a new building block of building block of building block of our platform. Therefore, our platform. Therefore, our platform. Therefore, it very quickly it very quickly it very quickly opens up new opens up new opens up new opportunities for

opportunities for opportunities for implementing implementing implementing agent systems at Monzo. agent systems at Monzo. Now we can Now we can safely connect safely connect safely connect it with it with it with all our all our all our security, backend, and security, backend, and security, backend, and platform requirements in mind at the same time. platform requirements in mind at the same time. platform requirements in mind at the same time. Our platform Our platform Our platform is by definition is by definition is by definition independent of independent of cloud solution providers. So, cloud solution providers. So, if we go back if we go back if we go back to the diagram, the little to the diagram, the little to the diagram, the little blue-green blue-green blue-green agent block can easily agent block can easily agent block can easily be replaced with any be replaced with any be replaced with any model. We can model. We can model. We can dynamically choose dynamically choose dynamically choose the model that best the model that best the model that best suits our suits our suits our goals. So we are by goals. So we are by goals. So we are by definition not definition not definition not tied to any tied to any tied to any particular particular particular supplier. This gives supplier. This gives supplier. This gives us the flexibility to us the flexibility to us the flexibility to understand which ones understand which ones understand which ones best best best meet meet meet Monzo's needs. And finally Monzo's needs. And finally , we continue to , we continue to , we continue to move towards move towards move towards covering a wider covering a wider covering a wider development window. development window. Even if we Even if we stop development stop development stop development today, we today, we today, we can already cover can already cover can already cover much more: we are much more: we are much more: we are preparing changes in the preparing changes in the preparing changes in the backend, preparing backend, preparing backend, preparing changes in the application. changes in the application. We can test We can test these changes in the these changes in the these changes in the deployment environment and deployment environment and deployment environment and cover the cover the cover the data tests. Who knows?

data tests. Who knows? data tests. Who knows? Perhaps even Perhaps even Perhaps even predictors from the predictors from the predictors from the other side. So when other side. So when other side. So when remote agents remote agents remote agents can do this from can do this from can do this from start to finish, we start to finish, we start to finish, we can can can understand and understand and understand and integrate our integrate our integrate our systems much better. Again, systems much better. Again, systems much better. Again, the market is moving, it the market is moving, it the market is moving, it doesn't mean we're doesn't mean we're doesn't mean we're necessarily tied necessarily tied necessarily tied to agent to agent to agent technology forever technology forever . But these and other . But these and other . But these and other aspects make us aspects make us aspects make us believe that this is a believe that this is a believe that this is a very important part of very important part of very important part of our internal our internal our internal platform now. And the platform now. And the platform now. And the main takeaway main takeaway main takeaway I want you to I want you to I want you to make is: how you can make is: how you can make is: how you can benefit from an benefit from an benefit from an LLM, but not from LLM, but not from LLM, but not from traditional LLMs, but from traditional LLMs, but from traditional LLMs, but from data multipliers that data multipliers that data multipliers that you and your team you and your team you and your team can use can use can use to drive to drive to drive value. Thank you.