CEO iterate.com, prev. cofounder/CTO Monzo, dad of three
- Tweets 5,018
- Following 2,729
- Followers 8,346
- Likes 5,576
Pinned Tweet
2025 will be the year we see the first self-driving startups.
Level 0: No AI
People do everything. They come up with ideas, build products, and run operations. Many legacy businesses still work this way.
Level 1: People use AI tools ⬅︎ we are here
People might use ChatGPT to help write copy or Cursor to help write code. This is where most startups are today.
Level 2: AI agents complete tasks based on human instructions
People might ask AI agents to write software from a plain-English spec or tell it execute well-defined customer service processes. At this point entire departments (like support or QA) get largely replaced by AI. No startups I know of operate at this level yet—but if yours does, let me know.
Level 3: AI agents propose changes to their own instructions
They might propose new customer service processes and product changes in response to customer feedback. Humans would still approve each of those changes. Just a few people could run a large company this way.
Level 4: AI agents autonomously change their instructions
At this point startups become self-improving. Humans would only be involved as an escalation point or where required by the real world (e.g. to raise capital or to incorporate). At this point many startups would only have one human.
Level 5: No humans
AI agents decide which businesses to start, raise capital (through crypto tokens or other means), build and run them. No humans required. This would require major reforms in the legal and financial system.
19
21
218
76,219
After almost 9 years, today is my last day at @monzo. I've written a bit about it here:
monzo.com/blog/thank-you-jon…
🩷🧡💛💚💙💜
30
8
235
70,079
In a few years customers will expect all their transactions to be enriched in this way. We call it transaction “unfurling”
Having a little bit too much fun with Monzo's new receipts API. (❤️ to @kieranmch and the team at @monzo cc @t_blom)
Uses Gmail APIs to fetch emails, some logic to scrape known receipt formats & matching behaviour to push onto a Monzo TX.
4
24
147
I'm pleased to announce that this little nugget arrived safely a few weeks ago and we're totally smitten! @leahtemplewoman insists I'm the father but I'm not so sure... 👶
16
1
127
In my mind LLM "agents" are just event-driven programs (or as @karpathy would say "Software 3.0" programs)
1) They can receive new events (e.g. a "user message")
2) Some events trigger a "computation" (i.e. an LLM request)
3) This produces additional events (e.g. an "assistant message") that other actors can subscribe to
Under the hood the agents maintain an append-only immutable log and a reducer function that turns the event log into usable state - most importantly the input args for an LLM request (model options, system prompt, message history, tools)
Modelling agents this way makes it easy to implement all manner of fun asynchronous interactions. Handoffs to other agents. Async tool calls (that complete in the far future). Scheduled events. Consuming webhooks / messages from third parties etc. And it's easy to do forking, snapshotting, time travel, debugging etc. And normal techniques (e.g. debouncing) from event driven programming (i.e. Javascript) are then easy to use, too.
One thing I've recently been thinking is that you could even use a reactive querying mechanism like livestore or convex and give the agent a tool to query data. And then when the data changes the agent gets a new event.
Are there any existing agent libraries that encourage this design pattern?
20
6
133
34,394
Note to self: don’t go on a trip to SF when wifey is 35 weeks pregnant - even if your other two kids arrived after 42 weeks 🤣
Thankfully I made it back in time for the birth and bubba no3 arrived safe and sound 👶🥰
20
129
5,844
What if Cloudflare bought bun and merged it with workerd? I realise this makes little technical sense, but as user I would love bun’s DX on cloudflare’s platform and I’d love to be able to use the same runtime my deployed code uses when writing scripts or in my test runner etc
9
2
120
23,939
📢 I am delighted to announce that my longtime mentor and friend @Geek_Manager will be joining @monzo as CTO! I am super excited to finally get to work with her 🙌
I will take a broader executive role focusing on organisation design 🚀
4
5
105
I'm so proud of the team for this. A few people got together and built this in their "Monzo time", which is time that is set aside for working on whatever you want monzo.com/blog/2018/05/16/ga…
5
26
99
I was in a WhatsApp group called “Fintech AI founders” in London and nobody in it understood this 🤦
13
85
14,671
👋 Hey folks, I'm looking for a chief of staff to help run Monzo's tech org. This is a great chance to get access to and influence the decision making in a growth stage, pre-IPO tech company. The job ad is here - I would appreciate any intros 🙏💖 DMs open boards.greenhouse.io/monzo/j…
3
23
84
This is absolutely blowing my mind right now medium.com/@connorgettel/mas…
4
18
84
The thing that makes doctors and lawyers different is that it’s harder to verify the work is correct
Certain kinds of engineering work have high recall, high precision validators (e.g. competitive programming puzzles or refactoring a well tested codebase)
In those cases, as soon as the AI can get it right sometimes, you can just insert more money until the AI gets it right however often you want. That’s why o3 is ranked 175th on codeforces
Much harder to do with legal work
5
1
81
26,370
If you like distraction-free hacking with smart people in the beautiful english countryside, apply here!
The first step of my masterplan to build a thriving tech ecosystem in the cotswolds 🇬🇧 🌄
11
6
79
6,455
Over the last month I've moved from chatgpt to claude for deep research and then again to @p0
Each time my results got massively better, and as a result the number of times I used the service has increased
I know parallel.ai is meant to be an API service first, but I just use their playground UI for all my research now. It makes it way easier to export a clean markdown (or json!) of the results and, crucially, I'm not limited by some invisible deep research limit
6
8
65
14,432
These comments by @mitchellh capture my own feelings towards hard work in startups. At Monzo I didn’t feel like we could openly talk about that, but it was absolutely a large part of our early success
3
2
61
10,356
Super interesting presentation by @jasonmaude from @StarlingBank on building a reliable core banking system. I’m really glad that we can now have this level of exchange of ideas in the banking industry :) infoq.com/presentations/clou…
2
9
52
It’s still next token prediction
But in order to assign the probability of each next token, some clever neural pathways get activated. Just like with humans
1
1
48
7,663
This here is the right model. We need something like this in the UK that CS students can do instead of an internship at a hedge fund
Students - spend your summer building with @ycombinator
We're offering equity-free grants of $20k plus $90k in compute credits.
ycombinator.com/blog/summer-…
2
2
45
4,384
I asked 100 CS students at Imperial in London the same question yesterday. Less than 10 hands went up.
This is amongst students that had self selected into a talk about AI in software engineering.
The productivity gap between AI-pilled and normies will continue to widen
At stanford today I asked the cs 153 class how many of them use @cursor_ai. 300 hands shot up. Every. Single. One.
7
44
4,370
I noticed using AI makes people better at leading real people. It builds the muscle of saying exactly what you want, delegating and giving crisp feedback etc. Especially useful for young / inexperienced leaders
2
5
42
9,124
If you
1) are an extraordinary programmer
2) put in the work to get good at using AI
you can can build anything
Most extraordinary programmers I know still don’t want to do 2)
It’s innovators dilemma. They are so good at their job and so far ahead that the new thing actually holds them back for a while. Until it doesn’t.
I've shared the full transcript of every agentic coding session from implementing the unobtrusive Ghostty updates and provided commentary alongside about my thinking and process. Total cost: $15.98 over 16 sessions. "Vibing a Non-Trivial Ghostty Feature" mitchellh.com/writing/non-tr…
2
2
43
5,889
Made a script at a hackathon yesterday that lets me control my computer with my voice better
I can write typescript functions that intercept superwhisper transcriptions and then activate based on "spells" (like "raycast" or "open file" when in cursor). These spells can do anything!
Would be cool if somebody else could test it!
5
2
42
3,287
I’ve never been more excited about anything we’ve built at Monzo. Massive thank you and kudos to everyone across the company who’s made this a success. Try it and let us know what you think :)
36
I really love cloudflare’s tech and am excited to build on the platform. Espefially @KentonVarda ‘s new dynamic workers and DO facets
But I do have to agree with @cramforce - The whole “region earth” and “run code at the edge” thing doesn’t help my database application.
It made sense to build out the edge network this way when CF was a global caching and security layer
But for normal application use cases, I would prefer if CF just built a huge DC where I want it. I’d pay more for that, too
6
43
6,458
The way I use my computer has changed more in the last 2 years than in the 30 years before that. I think even more than it changed when I got a smartphone, because I just use a smartphone like a small normal computer
I now speak instead of typing where possible, I have at least 10 in-depth conversations about some technical problem with various LLM models and I use deep (re)search multiple times a day to search hundreds of websites to pull out just the info I need
I think the biggest difference is not the modality of input (speech).
It’s two things
- The computer can give input into my process. It brings stuff that I don’t have to initiate or know even exists (I just need to know how to ask open questions)
- I can give _imprecise_ input and still get something valuable
It’s pretty wild
3
2
41
4,698
Just got access to GPT builder and made this Elixir language assistant: chat.openai.com/g/g-dNnJ2FhT…
The only data I uploaded to it so far is a JSON dump of the Elixir core 1.16-rc.1 hexdocs
Let me know how it fares!
#MyElixirStatus
2
7
37
4,994
📣 Today we launched a fun side project
Go to garple.com to buy a fun domain like balses.com, exagi.com or flarse.com for your project or startup
We've got over 100 domains and they are all short, easy to say, meaningless and easy to trademark. The perfect blank canvas. Just like google, twilio, quora or our own company nustom.
4
5
37
12,564
It's been a year of spending most of my time coding again, after having mostly ascended into bureaucracy for a few years. Time flies when you're having fun!
5
38
1,801
I realise this is only relevant to a super niche audience, but nonetheless: here's our CLI tool guiding cursor through the steps necessary to build an entire app.
Again, the only instruction in the .cursorrules file is "Run `bifbof next` until it succeeds". And Devin and windsurf can also get through this process all the way to the end
4
38
12,745
I’m super proud of our customer operations team for this. They care a lot and it shows in rankings like these 🥰
It's official: we're the best bank for Overall Service Quality and Online and Mobile Banking Services 🎉
We also came fourth in the Overdraft Services category!
monzo.com/blog/2020/08/17/we…
1
26
Had the most productive coding day of my life. Yesterday I made a good .cursorrules file for a greenfield project using vanilla tech that LLMs know well (typescript, zod, vitest, trpc, etc)
Using cursor agent in new YOLO mode locally and a couple of devins remotely. They are all instructed to
1) Run until `pnpm test` passes
2) Add to .cursorrules
I thought the coding influencers were exaggerating, but this actually works super well. Even for reasonably complex systems engineering that would be beyond most junior engineers
2
34
2,852
I want to meet London-based AI-maximalist startup founders and set them up with our AI chief of staff
Ideal time to apply is just after you raised a seed round and are looking to "professionalise" (i.e. no bank accounts, accounting, HR setup yet)
DMs open - next onboarding event is July 15
ℹ️ We’re looking for 2–3 users to trial our operating system for autonomous startups
DMs are open if you're an early-stage, AI-maximalist founder based in London who’s recently raised your first round — or is in the middle of one
3
8
33
5,306
And now, back to work 🙇♂️
We're excited to announce that we've closed a new round of funding! 🎉
monzo.com/blog/2018/10/31/bi…
3
1
29
The bigger problem is that the got 5 rollout on the responses API was riddled with bugs and undocumented edge cases
For example, it is no longer possible to use built-in tools when using tool_choice=required. Why?
The complicated linking of reasoning items to subsequent items is v fiddly and poorly documented. Why throw an error when I change the order of input items subtly? I think there is also a bug where if the bot returns a reasoning item after certain built in tool calls, then the subsequent request of your agent loop will always fail saying the reasoning item was submitted without its corresponding output items
Why is it no longer possible to use the image generation or code interpreter tool with minimal reasoning?
7
33
13,904
This but in the British countryside
This is insane:
The world's best hackerhouse just opened in Finland, Europe 🇫🇮🔥🇪🇺
FR8
It's a literal hotel – 100 beds
Live there for free, with some of the smartest kids you'll ever meet, work on your most ambitious projects – every day. 🔥
Hard to get in but link in reply.
3
30
2,529
I'm so excited for this! Monzo's mission is to make money work for everyone - but I've always been just as excited about seeing what companies ex-Monzonauts start and how it impacts the London tech industry monzonauts.com/
1
3
29
So this is the point in a startup where you "suddenly" realize that you need more managers, structure and training. It's kind of nice. 🤔💡🚀
2
30
How to get stuff done in a growing organisation:
Step 1: Commit publicly to getting it done
Step 2: It will happen
1
6
27
It made us build on them and they_are_ v cool
But DX is still incredibly rough. V hard to debug anything
Highest value features would be
1) debug UI that can list DOs and show logs - should also work in miniflare
2) OTEL tracing (for all workers RPC including to DO instances), logs auto-tagged by span ID, etc
Can all be optional to avoid perf penalty when not needed
3
30
2,302
Looking for experienced system/platform engineers with an interest in AI to join our founding team in the UK 🇬🇧
Building an application and workflow platform on Cloudflare in typescript
Email me directly jonas@iterate.com with links to some of your work
10
7
29
2,497
Super proud of @monzo for this: monzo.com/blog/2018/05/30/ge…
3
25
For example, here's all the code it takes to give the agent a tool to send send itself a reminder event in the future (or on a recurring basis)
Giving this tool to an LLM leads to incredible emergent capabilities. All just using human language for "programming". E.g.
- "When you think you've done the task, set yourself a 10 minute timer and close it if the user hasn't responded by then"
- "Check the website once a month to see if it's updated meaningfully. If it has, re-open this thread"
(Of course we're massively benefiting from using durable objects and cloudflare agents SDK here)
In my mind LLM "agents" are just event-driven programs (or as @karpathy would say "Software 3.0" programs)
1) They can receive new events (e.g. a "user message")
2) Some events trigger a "computation" (i.e. an LLM request)
3) This produces additional events (e.g. an "assistant message") that other actors can subscribe to
Under the hood the agents maintain an append-only immutable log and a reducer function that turns the event log into usable state - most importantly the input args for an LLM request (model options, system prompt, message history, tools)
Modelling agents this way makes it easy to implement all manner of fun asynchronous interactions. Handoffs to other agents. Async tool calls (that complete in the far future). Scheduled events. Consuming webhooks / messages from third parties etc. And it's easy to do forking, snapshotting, time travel, debugging etc. And normal techniques (e.g. debouncing) from event driven programming (i.e. Javascript) are then easy to use, too.
One thing I've recently been thinking is that you could even use a reactive querying mechanism like livestore or convex and give the agent a tool to query data. And then when the data changes the agent gets a new event.
Are there any existing agent libraries that encourage this design pattern?
3
2
27
4,267
Okay folks, garple is live on product hunt from midnight pacific time
Check it out here: producthunt.com/products/gar…
I can promise the most unhinged launch video you have ever seen, as well as a shocking plot twist in the future 🔮
Startups often struggle with naming. Early at @monzo, we had an ugly 12 month trademark dispute over “mondo”, our original name.
So @jonas, my cofounder at @monzo, started a side project that sells cheap .com domains that would make good startup names like wirps.com or flarse.com
They are unique, short, easy to say and don't mean anything, so you can fill them with meaning - just like Google or Twilio did.
Check it out at garple.com
3
4
25
8,337
In other news, we're pivoting away from code generation tools to just making all the startups ourselves: iterate.com
2
28
3,045
Thanks to @simonvc for reminding me of this oldie but goodie: the @monzo branch locator API at api.monzo.com/branches
3
27
💡 Fun fact: When you pay for a flight, the merchant can choose to attach metadata about the booking to the card payment authorisation. This reduces the amount of interchange they have to pay. Thanks to @dingwallr's Monzo Time project you can now see this data in the Monzo app 🙌 nitter.app/MakingMonzo/status/100…
2
4
25
What's better than a slackbot that can connect to MCP? A slackbot that can connect to an MCP registry!
Fast takeoff in 3, 2, 1
Soo we finally did the thing we have been saying we will be doing: blog.modelcontextprotocol.io…
4
3
27
3,470
👋@monzo is looking for Cassandra contributors to build a bank from scratch 🏦💜🚀 DM me if you're interested or know somebody who might be :)
2
15
25
🚀 Nustom Summer Internship 2025
Are you a CS student or otherwise competent coder that's interested in startups and AI? Come spend three months with us and launch a revenue generating startup!
The only catch is that you have to use our AI generator to build your startup from human language input. And if it’s not able to, you need to improve our generator so that it can.
Fill in the form to apply: nustom.notion.site/150622a33…
3
8
28
8,378
Looking for 1-2 AI maximalist typescript hackers to join an internal hackathon in London in person next Thursday evening to play with the first version of our operating system for software 3.0
5
2
26
2,081
This is a quality post from @obeattie . I'm really proud of the engineering culture we've built at @monzo :)
On our blog today, @obeattie shares some of the lessons and best practices we've picked up over the last three years building Monzo 💪
bit.ly/2lHCE3l
3
24
Thanks for sharing the profiles of these amazing folks 💜 They are some of the best people I've worked with 💯
One thing to bear in mind: this stage of process is about roles. If we had 10 people in a role and needed to reduce that to 9, all 10 people would be considered at risk
1
4
25
In the future people will work with AI agents to make software, even if they have no technical training.
How will humans and AI work together what will that collaboration space look like?
We're looking for engineers/designers to explore this
DMs open
3
2
22
3,764
Async coding agent maxxing guide:
- Create great documentation
- Use your coding agents to keep it up to date
- invest in verifiers - tests, testing harnesses, static analysis, critique model prompts, etc
- Make it easy to spin up development environments in different kinds of VMs / containers
- Get your gitops right
- good linear ops - spec all your tickets etc
- Don’t use bleeding edge or unpopular tech
- keep human language workflows for common tasks in repo (e.g. modify-app.md) . Ask agents to use filesystem to store task state
- Find a way to turn context into each agent’s preferred format (e.g. using vibe-rules)
- Choose architecture where agent mistakes have small blast radius
Boring stuff really
Goal should be to spend more money on inference than salaries in a few years. Headcount kills
2
4
26
2,543
The latest Steve Yegge piece should be required reading to anyone involved in software development
Ignore at your own peril
If you haven’t gotten much value out of coding agents or are skeptical still, please read it all (it’s entertaining, too!)
5
3
24
4,302
Super excited for Charles to join - amazing impact after only four days
Time from first DM convo to onboarded on slack: one week
This is the good thing about X! If you break through cringe barrier and post about your work, other people might respond saying they are working on the same thing and then you can work together 💪
5
24
4,573
Fine we’ll do it on Monday. At the moment this doesn’t really work yet so may be a train wreck. Will see 🤣
In any case, fun will be had! @isnit0 hat cursor whip up a lil video for the occasion
Shall we do "twitch plays pokemon" but for iterating on a startup?
Start with a "neutral" startup that does nothing. Then it changes itself based on your feedback
2
2
24
13,949
Has original Facebook platform vibes
Bring your api token, bring your GPTs, bring your memories, bring your tools etc
24
1,774
I read this and found myself mostly agreeing. I guess one counter point would be that the author is essentially saying that capital will be more directly convertible into value/impact. Whenever this happened in the past (e.g. when the internet or cloud computing came along), the number of startups started with little capital _increased_
The difference with AI might be that you can reliably 1000x the investment and get 1000x the result (e.g. by starting 1000 startups, all with identical top shelf AI founders)
2
1
20
5,889
Shall we do "twitch plays pokemon" but for iterating on a startup?
Start with a "neutral" startup that does nothing. Then it changes itself based on your feedback
6
1
21
7,457
Working well with regulators will be key to the long term success of finance startups - where people’s money is concerned, you can’t just do whatever you want and that is a good thing bloomberg.com/opinion/articl…
23
Amazing work from the whole @monzo team 🚀
Hard work and grit paid off for @monzo, surviving two tough years in '20-'21. With a mission to make money work for everyone and an NPS of 70+, their customer obsession saved the business.
Now serving 1 in 6 UK adults, Monzo is the 7th largest UK bank by customer numbers—all within 9 years of launch. And the story is just beginning.
Congrats to TS (@TSMonzo), Sujata (@skaladeen), and the Monzo team on an incredible turnaround that will be remembered in history!
22
2,609
So pretty 😍 when can we switch from normal RBS business bank accounts to this?
.@11FSTeam launched a proposition today for @RBS called Mettle - it's all about forward looking banking to help entrepreneurs make better decisions.
We're super excited as this is just the beginning. Come help us shape our product and get involved here mettle.co.uk
1
1
21
Watch our CLI tool bifbof send cursor agent on a 40 minute long quest to build doorwards from scratch (with a minor intervention to break it out of a loop)
I just wish we didn’t have to type “continue” all the time to overcome the tool calling limit @cursor_ai
Excellent work by @ofogelin who only joined part-time two weeks ago 🚀
3
22
2,731
Super cool to see @supabase publish prompt fragments for their tech. There needs to be a standardised way for everyone to make such context available to LLMs to use their technology or service
supabase.com/docs/guides/get…
2
22
4,005
Been waiting for this since the day we started. Super job @kieranmch and @simonvc 💪
IF you like doing cool things with your money
THEN you'll love our newest integration.
Introducing: Monzo on @ifttt bit.ly/2JqtP8v
1
1
21
Watch Claude code CLI use our bifbof cli to repeatedly get the next task in a long -ish workflow to generate an entire application 💥 (video sped up!)
Courtesy of @CFJWilliamson
2
21
1,286
TIL that Monzo is now the 7th most popular iPhone app in the UK
I wonder what percentage of current accounts opened in the UK are Monzo accounts now. We must be getting close to 50% now
Either way, what an incredible accomplishment by the @monzo team!
4
1
22
2,023
I’m sympathetic to your cause. But honest question: what are some examples of high impact projects that people didn’t work super hard on? There isn’t much evidence that not working super hard can increase odds of success of a project
I hope for everyone’s sake this changes - maybe with AI where people have more leverage
3
19
12,734
This talk is well worth a watch - especially if you’re concerned about either prompt injection attacks or how insanely inefficient agent loops with MCP tool calling are - where basic actions like “list some stuff and do something with each item in list” take way too many LLM round trips
It also seems vastly easier to think about sandboxing and permissions when dealing with TypeScript APIs instead of REST. I illustrated this in this talk I gave recently: piped.video/watch?v=xUj4HQt_…
5
22
6,585
Make something people want. Over 60% of @monzo users from the January 2016 cohort made a payment last week 📈🚀
1
19
💬 I'll be doing an AMA on @monzo's community forum tomorrow. If you have any questions about tech at Monzo, Monzo in general, our appointment of @Geek_Manager as CTO or just about myself, feel free to swing by: community.monzo.com/t/monzo-…
1
2
20
Wow, this competitive intelligence dashboard about challenger banks is blowing my mind: publicstats.itxdev.uk/dashbo… ❤️🙏📈 @youlikethaaaat
3
20
Here's a little exploration of how we can provide better guidance to coding agents like Cursor/Windsurf/Devin better
The core idea is that you could have a very small .cursorrules/.windsurfrules/etc file like this:
"After any task, run `bifbof boop` repeatedly and follow the instructions until it succeeds."
And then `bifbof boop` (for lack of a better name!) does all kinds of smart stuff to pull in the right context, run validators etc.. The community could even share bifbof plugins etc with each other.
Or maybe bifbof has multiple different commands and you just say in your cursorrules "run bifbof --help" to see all you can do in this repo" and there might be commands like "bifbof make-backend-endpoint" and "bifbof make-frontend-page" that guide the agent through the process of accomplishing the task.
A "SaaS boilerplate" would come with its own set of boilerplate specific bifbof plugins that could be used by any coding agent (or human!)
Is there any prior art on this?
3
1
21
2,989
Come to this! Good vibes and chaos guaranteed! There will be pizza
Last week @jonas hosted the UK’s first "World’s Shortest Hackathon" – just 3 hours to build something awesome with any AI tool. We proved it’s possible, and now we’ve teamed up with @join_ef to scale it up after work on Monday!
Apply now: london.aitinkerers.org/p/uk-…
3
21
1,548
Just used @WisprAI , @cursor and two joycons to write some code while walking on a treadmill 🤯 Feels like the future
I'm new to it so it's janky and I had weird audio issues, but I got the feature coded hands-free, which is amazing for my RSI!
1
1
19
2,400
Internship candidate from Cambridge CS is in a startup society that runs a competition where the first thing you do is write pitch and try and get funding. It is not recommend to start writing code before that.
It's completely backwards.
It feels like the whole system here is trying to nerd-snipe me into making something better.
6
20
1,485
Building on durable objects is so fun. Took me 10 LOC to make our cloudflare agents subclass automatically resume LLM requests that got cancelled because the durable object crashed or got evicted or whatever
1
1
21
2,121
Compiling shared libraries into services makes changing those libraries frustratingly difficult. We try to avoid this and turn the libraries into services instead. @obeattie I wonder if now is the time to write that blog post on this 🤔 nitter.app/segment/status/1016703…
1
1
20