How Monzo built an in-house coding agent that grew from a Slack emoji reaction into a system authoring over 10% of merged PRs and running more than 1,800 tasks a day. Felipe Felisola Caso and Francesco Galletta attribute its success to existing 'lots of little multipliers': an MCP gateway into internal tools and knowledge, sandboxed agent orchestration suited to a regulated bank, and a platform treating the agent as a building block — now used for incidents, on-call data quality, and requests from non-engineers. Chapters cover Monzo's scale, the viral Slack agent, the MCP gateway, sandboxed orchestration, incident mode, non-engineer contributions, and the rationale for building in-house. # Channel / Video Identity Observations **Video:** "A Bank Built Its Own Coding Agent. It Took Off" — Felipe Felisola Caso & Francesco Galletta, Monzo ## Speakers & Affiliation - Self-identified as **Monzo employees** presenting on their coding agent: > "We are Felipe and Francesco... what is Monzo? We are a bank. Digital bank trying to make money work for everyone." - The tagline *"trying to make money work for everyone"* is presented as Monzo's mission statement. ## Intended Audience - **Engineering conference audience**, apparently fellow practitioners — referenced repeatedly: *"you have heard a lot of success cases of companies adopting AI agents"*, "here in the conference." - Awareness of prior conference content: *"a conference a couple of weeks ago here in New York... this article came out"*, and *"probably everyone has heard in multiple sessions."* - The closing takeaway addresses builders directly: *"how can you benefit from LLMs... lots of data multipliers that you and your team can use."* ## Claimed Scope (Creator Claims) - **Scale of engineering estate:** *"over 3,000 back-end services, around 13,000 data warehouse models, and dozens... of ML predictors."* - **Agent adoption metrics:** Agent Chip *"authors over 10% of all merge PRs across tech"* and *"runs more than 1,800 tasks per day."* - **Core thesis (explicitly redefined branding):** *"LLMs in this case stands for lots of little multipliers"* — success attributed to a strong internal platform, not the model alone. - **Regulatory positioning:** *"we are a bank, we live in a regulated environment, so we want to make sure that the non-determinism from the agents is controlled."* - **Vendor neutrality claim:** *"Our platform is provider agnostic by definition... not specifically tied to any provider."* - Explicitly disclaims a bragging framing: *"Are we saying we want to brag that we built Claude Tag 6 months before Anthropic released it? We could be. That's not the focus."* ## Observable Presentation Evidence - **Timeline claims:** a Slack channel "went viral" in *"January 2026"*; the talk is framed within *"the state of the tech industry in 2026."* - **Concrete product behaviors described:** reacting with an emoji to summon the agent, code reviews ("looks good to me"), Linear ticket assignment, incident-mode investigation, "Agent Chip on Call," and non-engineer self-serve workflows. - **Architecture walkthrough demonstrated:** MCP gateway, sandboxed orchestration in an isolated namespace, controlled egress proxy, blue-green agent pod. - **Naming quirk:** the agent is called both "Agent Chip" and "Agent Ship/Agent Shit" in the transcript, suggesting either stenographer error or a deliberate joke ("little emoji here. Really cute."). ## Confidence & Limits - Evidence is drawn solely from one talk's transcript; no channel-level conclusions about broader catalog or recurring series are warranted from a single conference presentation. - Claims (PR percentages, viral adoption, provider agnosticism) are creator-stated, not independently verified. # Delivery Observation — Monzo "Agent Chip" Talk ## Speaker Setup - **Two presenters** ("we are Felipe and Francesco") splitting a co-presented talk; one speaker appears to deliver most of the transcript, with a reference to "the ones Francesco mentioned" and "as Francesco mentioned" indicating handoffs not fully visible in this transcript. - Opening is casual and **self-deprecating about tech failures**: *"Well, that was loud"* and, during a slide glitch, *"Oh, no, what happened? You cannot do this. It's all right."* ## Tone & Style - **Conversational, humorous framing of serious material.** A joking industry quote is flagged with *"hopefully jokingly quote"* and *"So, some food for thought, right?"* Humor also lands on product personality: *"just have a chat with you, but I don't think he likes it so much."* - **Rhetorical question → answer pattern** used repeatedly to move the talk forward: > *"but so what? Why are we What are we trying to share today? Are we saying that we want to brag that we built Claude Tag 6 months before Anthropic released it? We could be. That's not the focus of today's talk."* > *"But, what if the errors never happened in the first place?"* - **Anticipating the audience's thoughts** ("You can be thinking now, the market is moving fast...") is a recurring bridge between sections. ## Recurring Vocabulary & Metaphors - **"Building blocks"** is the dominant organizing metaphor, appearing at least four times: *"With these as building blocks, any use cases that come after it are much, much easier to be built on top of"* and *"Agent Chip being composable, it is now a new building block."* - **"Boxes" and nesting** for architecture: *"that box inside the box inside the box is where the nondeterminism happens."* - **Signature acronym gag, set up and paid off**: *"it's LLMs. But actually LLMs in this case stands for lots of little multipliers"* — later reworded to *"lots of data multipliers"* in the closing takeaway, suggesting the bit is intentionally repeated as a bookend. - **"Secret sauce"** framing for the core message: *"We want to share the secret sauce of what makes Agent Ship a successful product."* - Repeated stress words: **"safely"** (*"able to safely give it access to our data"*, *"the more data we are able to give safely"*), **"controlled"**, and **"non-determinism"** — consistent with the regulated-bank framing (*"we are a bank, we live in a regulated environment"*). ## Teaching Habits - **Concrete scale numbers up front** to size the problem: *"over 3,000 back-end services, around 13,000 um, data warehouse models."* - **Progressive diagram walkthrough**: the architecture is introduced, then "exploded": *"if we explode this to give you a bit more under the hood visibility"* — with color-coding called out explicitly (*"the little uh blue-green box there"*, *"I just use a different color to show this is the more than 3,000 services"*). - **Callback structure**: earlier content is re-referenced later (*"if we come back to the diagram"*, *"the ones Francesco mentioned"*). - **Enumerated takeaways** near the end, signaled with *"Um and finally, we are continued to move..."* and closed with *"the key takeaway that I want to I want to you all to take from is..."* ## Filler & Verbal Tics (observable in transcript) - Heavy use of **"uh/um"** throughout, plus **"So"** as a near-continuous sentence starter (*"So, we do all sorts of things"*, *"So, we have remote agents running..."*). - Some doubled/self-corrected phrasing typical of live delivery: *"the state the state of the tech industry"*, *"multiple of multiple providers"*. ## Confidence Notes - **Strong evidence** (multiple repetitions): "building blocks" metaphor, rhetorical-question transitions, "So" openers, the LLMs-redefinition gag. - **Single-instance observations** (flagged as such): the "secret sauce" line, the slide-glitch moment, and the humorous quote — these appear once each, so no pattern should be claimed from them. - Speaker attribution between the two presenters is **uncertain** from this transcript alone; most text appears to be one continuous speaker. # Observation Record: "A Bank Built Its Own Coding Agent. It Took Off" **Speakers:** Felipe Felisola Caso & Francesco Galletta (Monzo) — attribution per video title; transcript itself uses "Felipe and Francesco" without consistently distinguishing which speaker says which line. --- ## Context and Company Background - Monzo is described as a **digital bank** "trying to make money work for everyone." - Scale claims attributed to the speakers: - **Over 3,000 back-end services** maintained - **Around 13,000 data warehouse models** - **Dozens of ML predictors** (credit decisions, financial crime, app/backend/data use cases) --- ## Agent Chip: What It Is and Its Adoption Claims - In **January 2026**, a Slack channel mentioning Agent Chip "went viral" inside Monzo. Users trigger it by reacting to a message with an emoji, which acts as a prompt. - It evolved into an **"always available teammate"** in Slack: - Answering questions and architecture queries - Performing code reviews ("like a very senior engineer, you will just tell you it looks good to me") - Accepting assigned **Linear tickets** - Quantified claims stated by the speakers: - Authors **over 10% of all merge PRs across tech** (data, ML, backend, app) - Runs **more than 1,800 tasks per day** - Beginning adoption by **business users** (answering business questions, writing queries) --- ## Core Thesis: "LLMs = Lots of Little Multipliers" - The speakers explicitly reframe the acronym: **"LLMs in this case stands for lots of little multipliers."** (Note: in the closing, the speaker says "lots of data multipliers" — a slight inconsistency in wording that should be preserved.) - They argue Agent Chip's rapid success came from **Monzo's pre-existing platform capabilities**, not the agent itself: - Alerting, monitoring, third-party integrations - A staging environment for easy testing - Queuing abstractions, easy deployment, database abstractions - They explicitly disclaim a bragging motive: *"We could be"* bragging about building "Claude Tag [likely Claude Code] 6 months before Anthropic released it," *"That's not the focus of today's talk."* --- ## Architecture Claims ### MCP Gateway (local-agent era, pre-Agent Chip) - Built first to serve **local coding agents**, exposing internal services and knowledge sources. - Purposes stated: **deterministic behavior** in choosing what to expose, plus managed **authorization/authentication** for different agents. - Rationale quoted: *"the more context and the more data we are able to give safely to our agents the better."* ### Agent Orchestration System (remote agents) - Powers Agent Chip: agents run in a **sandbox within an isolated namespace**, separate from production infrastructure. - Stated motivation: *"we are a bank, we live in a regulated environment, so we want to make sure that the non-determinism from the agents is controlled"* while still allowing a space to "run loose and develop." - Sandbox components described: - **Agent image** — defines exact tools, environment, CLIs, auth (via the gateway) - **Proxy** — controlled egress only ("we only get the egress that we choose to give it") - **Agent box** — where "the nondeterminism happens" ### Composability - Agent Chip itself became a **building block**: Slack, GitHub, Linear, and the 3,000+ services connect to it, and other parts of the bank build new use cases on top. --- ## Reported Use Cases and Outcomes - **Incident mode:** Agent Chip joins incident channels, investigates logs/traces/code. The speakers stress a deliberate constraint: *"We don't want the agent to be saying a lot of things during an incident... we want it to maintain high signal,"* with humans controlling how much it participates. - **Ask channels:** squad-to-squad queries; Agent Chip drafts responses; sometimes it resolves the ticket alone, sometimes it tags a human for more context. - **Agent Chip on Call:** reactive triage of data-quality alerts with automatic investigation; plus a **proactive version** monitoring backend changes for potential data impacts *before* batch processing. - Claimed outcomes: - Higher breadth/volume of tasks (engineers delegate to remote agents rather than "babysitting" local ones) - Automation of routine processes so humans focus on "higher value problems where the ambiguity lies" - **Lower barrier for non-specialists** to contribute: they can get tested drafts from Agent Chip before involving an engineer. Qualification stated: *"this doesn't change the rigor to which we develop software at Monzo."* --- ## Longevity Argument (Why It Stays Valuable in 2026+) - **Composability:** Agent Chip is now a platform building block enabling new agentic use cases with security/platform requirements met. - **Provider agnosticism:** the agent box "can be very easily swapped by any model"; they are *"not specifically tied to any provider by definition."* - **Broader development window:** drafting backend/app changes, testing in staging, data tests, "who knows? Maybe even predictors." - Explicit hedge: *"doesn't mean that we are necessarily married to Agent Chip until the end"* — even if development stopped, it would "already be valuable." --- ## Notable Tensions / Qualifications Preserved - The "Claude Tag before Anthropic" remark is offered as a possible brag but explicitly set aside. - Speaker 2's acknowledgment that a jokingly quoted X post ("if you're not building your own AI coding harness, are you even a serious tech company?") frames the industry zeitgeist — presented as "food for thought," not an endorsed position. - Non-determinism is treated as a risk to be contained (regulated environment), not eliminated. - Inconsistent acronym expansion: "lots of little multipliers" vs. "lots of data multipliers" in the closing. # Speaker & Provenance Record ## Video Metadata - **Title:** *A Bank Built Its Own Coding Agent. It Took Off — Felipe Felisola Caso & Francesco Galletta, Monzo* - **Video ID:** `01a15ca6-62b8-75bf-a2fb-704aa4cff91a` ## Speakers & Roles | Speaker | Role | Basis | |---|---|---| | **Felipe Felisola Caso** | Co-presenter, conference talk | Named in title and self-identification in transcript ("we are Felipe and Francesco") | | **Francesco Galletta** | Co-presenter, conference talk | Named in title; referenced by first name in transcript ("the ones Francesco mentioned") | - **Role type:** Both are identified as **Monzo employees presenting their own company's internal engineering work** at what appears to be a technical conference. - Audience applause is noted at the end; **no Q&A or audience speech is captured** in the transcript. ## Provenance Facts ### Company & Scale Claims (by the speakers, about Monzo) - Monzo described as a digital bank with **3,000+ backend services**, ~**13,000 data warehouse models**, and dozens of ML predictors. - Agent Chip, per the speakers, authors **over 10% of all merge PRs** across tech and runs **more than 1,800 tasks per day**. - Timeline claims: a **Slack channel went viral in January 2026**; talk framed as occurring "at this point in 2026." ### Third-Party / Cited Claims - Reference to an **article from a New York conference "a couple of weeks ago"** about the state of the tech industry in 2026, containing a quote (attributed only to "X") — *"if you're not building your own AI coding harness, are you even a serious tech company?"* - References to **Anthropic** and **"Claude Tag"** (as rendered in transcript) released after Monzo's internal tool; speakers claim Monzo built a comparable capability ~6 months earlier. - Repeated references to what attendees have "heard in multiple sessions" at the conference — context, not a specific citation. ### Tools & Integrations Mentioned - **Slack**, **GitHub**, **Linear**, an internal **MCP gateway/server**, an **agent orchestration system**, sandboxed isolated namespaces, and provider-agnostic model swapping. ## Promotional Segments - ⚠️ **The entire talk is effectively promotional for Monzo's internal engineering platform and Agent Chip.** Speakers explicitly frame the goal as sharing "the secret sauce" while acknowledging the bragging temptation ("We could be. That's not the focus of today's talk."). - Metrics (10% of PRs, 1,800 tasks/day) are **self-reported by the presenting company** and not independently verified within the transcript. ## Availability & Limitations - No curated metadata beyond the title was supplied. - No comment section availability stated. - Slides are described verbally (diagrams, emoji, Slack screenshots) but **not visible in the transcript**; visual claims are inferred from narration only. - No external sources, links, or citations are provided within the transcript itself. \>\> Uh, we are Well, that was loud. Uh, we are Felipe and Francesco. Uh, we come to talk to you about the compound value of building Monzo sync house agents using LLMs. Um, so what is Monzo? We are a bank. Digital bank trying to make money work for everyone. Uh, to give you a size of the problem for coding agents, uh, we have our teams of engineers maintaining over 3,000 back-end services, around 13,000 um, data warehouse models, and dozens and dozens of ML uh, predictors. So, we do all sorts of things like deciding on credit, deciding on financial crime, and doing very complex use cases both for the app, the back-end, and data. Um So, I'm pretty sure by even by today, but outside of the conference, you have heard a lot of uh, success cases of companies adopting AI agents for coding. Uh, in fact, uh, there was a conference a couple of weeks ago here in New York, and this article came out talking about the state the state of the tech industry in 2026. Uh, and there is this hopefully jokingly quote from X saying that if you're not building your own AI coding harness, are you even a serious tech company? So, some food for thought, right? Uh, then inside of Monzo in January 20 26, this a Slack channel went viral. Agent chip. And hopefully I can Oh, no, what happened? You cannot do this. It's all right. So, hopefully I can bring your attention to this little emoji here. Really cute. We love it. Uh, this became really, really viral inside of the company for a really good reason. You can give it uh, any prompt directly by just reacting with the little emoji to the message and it will pick it up as a, you know, prompt. Uh, so that evolved, we double down, and that became an always available teammate inside our Slack and whole environment. Uh, to give you a few examples, it does feel like a coworker. Uh, it's just there. You can always at Agent Chip and get it to do something for you, answer a question, find out something about some architecture, or just have a chat with you, but I don't I don't think he likes it so much. Uh, it can also do code reviews. So, like a very senior, uh, engineer, you will just tell you it looks good to me and go on your merry way. And you can also assign, uh, linear tasks, as well. So, you can give it tickets and it will just get into it. Um, so all of this is good and well. Uh, forward into today, Agent Chip authors over 10% of all merge PRs across tech, so that's data, uh, machine learning, that's, uh, back in engineering, app, etc. Also runs more than 1,800, uh, tasks per day, which is a lot. Um, and it's being actively, uh, starting to be adopted by business users as well. So, it's not just coding and tech, it's also trying to answer business questions, writing queries, or self-serving anything that a business user may need. So, again, you've heard many success stories about companies adopting AI coding agents, but so what? Why are we What are we trying to share today? Are we saying that we want to brag that we built Claude Tag 6 months before Anthropic released it? We could be. That's not the focus of today's talk. Um, we want to share the secret sauce of what makes Agent Ship a successful product and it's LLMs. But actually LLMs in this case stands for lots of little multipliers. We believe that what enabled us to build Agent Ship so quickly and so invaluable way is a very strong platform at Monzo. So when we say that this went viral and you could write a prompt, react, and profit. Actually this comes from a number of capabilities we already had. So alerting, monitoring, integrations with third parties such as the ones Francesco mentioned, a staging environment that allows us to test easily, queuing abstractions, easy deployment, database abstractions, and a number of other things I could be mentioning here. So first and foremost what we have in our platform is what allowed Agent Ship to be successful. But this does not stop there. We saw even before Agent Ship as a remote agent existed, we started to invest in supporting AI coding agents even in local environments. So the one of the first things we did was develop an MCP gateway that allowed easy access to a set of knowledge sources and internal tools and some of the services that Francesco mentioned. So this is roughly how the architecture looks like. We have the things we want to connect as the internal services and knowledge sources. And we put that behind the MCP server that we built to get the deterministic behavior choosing what are the things we want to expose to the agents and then we manage also authorization, authentication for the number of different agents. Again, this started in the world of local coding agents that wanted to connect and by by now probably everyone has heard in multiple sessions that the more context and the more data we are able to give safely to our agents the better. After that, we built the agent orchestration system, which is exactly what powers Agent Ship behind the scenes and makes it work in a sandbox and with very strict tools and permissions. So, if we go back to our diagram, the whole architecture for the MCP server basically condenses into this box and this is where we connect the rest of the Agent Ship architecture. So, we have remote agents running inside a sandbox in an isolated namespace and then the orchestration system triggers and then reads and gets this information from from that. The main idea here is and again, this is coming across in a multiple of multiple providers and labs and so on is that we want to separate this from the rest of our infrastructure in production. Again, we are a bank, we live in a regulated environment, so we want to make sure that the non-determinism from the agents is controlled. But at the same time, we want to have an environment where we can let it run loose and develop. And again, if we explode this to give you a bit more under the hood visibility of what is going on. So, in this isolated namespace, multiple pods come up with sandboxes that the agents are running on. The agent image is where we define the exact tools the agents will have as as access. So, things like the environment, the CLIs and the authentication and the authorization that again, we get from the gateway itself. The proxy is what allows us to connect to both the exterior world. So, that's how for instance Agent Chip is able to connect to our code, to our repositories, and work with that. Um but, of course, we only get the egress that we choose to do choose to give it. Um so, that is how we make sure that \[snorts\] it's very controlled and deliberately giving access to the things we want it to be able to see. And finally, the little uh blue-green box there is where the agent itself is running. So, that box inside the box inside the box is where the nondeterminism happens, and it's where the agent can perform everything that it wants. So, this is how the whole diagram looks like when we are able to give our MCP server access to everything that we care about for the agent, we are able to safely give it access to our data, and the agent orchestration system handles everything that's around it. With these as building blocks, any use cases that come after it are much, much easier to be built on top of. So, as Francesco mentioned, this we can connect Slack to, and then we can connect GitHub, and then we connect Linear, and then I just use a different color to show this is the more than 3,000 services that we can be connecting to to this as a building block. So, in the same fashion that we had multiple things in our platform that enabled us to build Agent Chip and get value out of it. Now that Agent Chip is also a building block, the multiple parts of our bank can also leverage it to create other use cases. So, we start to see this building blocks enabled us to not only get more um product features internally, but to drive concrete business outcomes. Um we observed a higher breadth and volume of tasks that engineers are able to run, as you you expect, because then we don't need people to be babysitting their agents all the time. They can use their coding assistant for um specific tasks, but they can also delegate to remote agents. We can automate routine processes much better. So, that allows humans to focus on the higher value problems where the ambiguity lies particularly. Some examples of these that we are able to do. Um this is Agent Chip working in incident mode and leveraging the MCP to get very relevant data. So, what happens here is whenever we identify and trigger a new incident, Agent Chip joins the channel and is able to start an investigation, understanding logs, understanding traces, being able to connect to that exact service, and of course, the code itself to see what's going on and to help the human with a very high signal. We don't want the agent to be saying a lot of things during an incident and testing hypotheses like we want it to maintain high signal, and we give the humans also control of how much Agent Chip will be able to support there. We can continue to use it on Slack, but not only by triggering it by mentioning it, uh but also in what we call ask channels. It's basically squad-to-squad interaction inside Mondzo. A squad needs to have some context on a back-end service, on a product development, on the road map, and Agent Chip can also, since it can connect to our knowledge sources, it's able to draft a response for the human in the loop to get as reference and support there. Maybe the response will be enough, and Agent Chip can just solve that ask ticket by itself. Sometimes it will ask for more context, and the human in the loop can be tagged um and come in to support more. Um something that we're doing very recently that's been uh creating a lot of value is also triaging data quality alerts and checking for back-end changes that can create problems. So, basically we call it Agent Chip on Call and it observes the failures that we have and triggers an investigation automatically. This is the reactive version of it. But, what if the errors never happened in the first place? We created based on this the proactive version that monitors back-end changes and identifies which are the potential outcomes that that can have for our data even before the data is processed in the next batch execution. So, again, all of these are examples of how our technology team is using it, but even better than that, we've observed a lower barrier for non-specialists in the company to contribute to our systems. With a very quick entry point for Agent Chip on Slack, non-specialists can, based on product feedback, based on ideas, based on discussions they are having, just call Agent Chip, get a draft that works, that is already tested in staging, and then, and only then, involve an engineer to make sure we get the changes reviewed and deployed properly. So, this doesn't change the rigor to which we develop software at Monzo, but this does allow a number of non-specialists to also start contributing more closely to the team. You can be thinking now, the market is moving fast. We are hearing a lot of exciting developments and new products and new features coming from multiple companies here in the conference and otherwise. What makes us believe that Agent Chip continues to be a valuable thing at this point in 2026 and moving forward? Um As I said and I like to to reiterate, Agent Chip being composable, it is now a new building block that we have in our platform. So, it very quickly enables new use cases for agentic implementation at Monzo. Now that we can connect it very safely with all our security requirements and all our back end and platform requirements at the same time. Our platform is plat provider agnostic by definition. So, if we come back to the diagram, the little blue/green agent box there can be very easily swapped by any for any model. We can dynamically choose which model best fits the purposes that we have. So, we are not specifically tied to any provider by definition and that gives us flexibility to understand which ones will fit Monzo's needs the best. Um and finally, we are continued to move to cover a bigger development window. So, if we even if we stop development today, this would already be valuable, but we can cover a much broader sense in terms of we draft changes in the back end, we draft changes in the app. We can test these changes in a staging environment, cover the data tests. Who knows? Maybe even predictors on the other side. So, when remote agents are able to do this end-to-end, we can understand and integrate our systems much better. So again, the market is moving, doesn't mean that we are necessarily married to agent chip until until the end, but these and other aspects of it make us strongly believe that is it is a very strong part of our internal platform now. And the key takeaway that I want to I want to you all to take from is how can you benefit from LLMs, but not the traditional LLMs, but the lots of data multipliers that you and your team can use to leverage value. Thank you. \>\> \[applause\]