Andrew Somervell
When engineers get 10x faster, the team becomes the bottleneck
In this interview, Andrew Somervell, co-founder of Hamster, explains how Taskmaster, the open source project he and his co-founder started in 2025, invented plan mode for coding agents and grew to 1.8 million downloads, and why the requests to take that acceleration to work became Hamster, a multiplayer AI workspace where product, design, and engineering agree on a brief before agents build it. He describes a ten-person team spread across San Francisco, New Zealand, Montreal, and Bangkok, why he keeps it small and composable, and how weekly direction and discovery meetings replaced quarterly roadmaps. He also talks through the launch that got a shrug, the grit it takes to keep iterating, and why he expects token prices to fall a hundredfold while harnesses, not models, win.

Andrew Somervell is co-founder of Hamster, a multiplayer AI workspace where product, design, and engineering agree on a brief before coding agents build it.
Andrew ran a software consultancy for ten years, placing engineering teams inside banks and other regulated enterprises and presenting at board level. In February 2025 he and his co-founder started Taskmaster, an open source project that breaks a product requirements document into tasks for AI agents; it has passed 1.8 million downloads and 28,000 GitHub stars, and Twitch founder Justin Kan became Hamster's first seed investor. Hamster's ten-person team spans San Francisco, New Zealand, Montreal, and Bangkok.
“Collaboration, trust, autonomy, agency. They're just key pillars.”
Andrew's core argument is that once engineers get a 10x speedup, the bottleneck moves to the team, so the work is agreeing on direction up front and giving people the autonomy to deliver.
"Hamster runs the wheel for you," Andrew explains. "It does the busy work so you don't have to."
The conversation
In this conversation: Jacqueline Samira (CEO, Howdy), Josh Rubin (Host, CTO Studio), and Andrew Somervell (Co-founder, Hamster).
I'm Andrew Somervell. I'm the co-founder of Hamster. Hamster is multiplayer AI for product engineering. In a world where engineers have gotten a 10x improvement in velocity, the new bottleneck is interacting with the team and going fast as a team together. Hamster aims to address that.
It's so true. I had this conversation with my team a couple of weeks ago, where for the first time engineering was no longer our bottleneck and product was. A lot of foresight for you to observe that. How long ago did you start the company? When did you realize this was going to be a bottleneck?
In February last year, my co-founder and I started an open source project called Taskmaster. It took a PRD and broke it down into tasks and subtasks, and back in the GPT-3 days that meant the AI could go sequentially through the project plan and deliver it piece by piece without blowing its context window or going off the rails. Fast forward to now, it has done 1.8 million downloads and 28,000 stars on GitHub. It's a great open source project, and we basically invented plan mode. By late last year plan mode became prevalent across all the harnesses, and then we had our Opus 4.5 moment in December and agents got pretty autonomous. But earlier than that, a lot of people were coming to us saying, how do I take this to work? How do I take the acceleration I've gotten as an engineer on my local machine and use it in a team setting? That was the kernel of the idea: build a SaaS tool that brings the local engineering environment into something you can work on with your team, engineering, product, design, and founders, to agree on a brief up front and align on it. Then it's easy to take that and ship it in your local harness or a cloud agent. And you've got an artifact to review whether what was done is what the team agreed. That's the bottleneck: you can ship so much, but if you're shipping in the wrong direction and the wrong stuff, it comes back into the team. Where did this come from? Once you've got that upfront agreement, it's easy to get the team to give a thumbs up, it gets delivered quickly, and it ships to production quickly.
How do you get that upfront agreement? Are you transcribing it all and then the AI puts it together?
Group chat and group documentation. You're seeing a few people figure that out now. Think Slack or iMessage, but with a collaborative document like Google Docs next to it, and the AI in there with you. We had to train it when to talk and when to stay out of conversations, but it can write the documentation with you as you're talking as a team. That's a really big unlock.
Do you also take in-person meeting notes? How do you get data that's in real life versus online?
We have 1,200 connections to other platforms, including Granola and Google Meet. We also have a Hamster agent that joins our meetings and listens, and can talk back. It has full knowledge of your business and everything that's gone before. You can ask it questions about your codebase. The knowledge graph and workflow graph underpin everything: you set direction, you do discovery together, you take a hunch or a rough idea, reason through it as a team, and then you deliver. That's it in a nutshell.
Talk to me about your go-to-market. How are people finding you? Was it through your open source project?
A large proportion of our website traffic is still the open source community, and a lot of X activity, founder-led: here's the latest process we've been able to automate and don't have to do anymore. The idea behind the Hamster brand is that it runs the wheel for you. It does the busy work so you don't have to.
I was imagining that sometimes when you're building, you feel like you're on a hamster wheel. I feel like we've been here before.
We didn't want to give people the impression they're on the hamster wheel. Hamster runs the wheel for you. There's a lot of work about the work, and a lot of agents out there will hand you something to do, or previous-generation task management platforms will help you write a status update. Why am I writing a status update, even AI-assisted, when AI can just do it? So what it lets product managers and engineering managers do is collaborate in one space, reason through what to build and what the insights are, and set direction, without the dance of: I create a markdown file in my Claude session, check it into a git repo, send it to you on Slack, you pick it up, and the context is lost. That's the individual nature of things. Hamster brings it all together in one space.
How does the business model work? How do you do pricing?
Per seat at the moment. If you really ramp up with our cloud agents, there's per-token billing on that, but we're trying to dial it in on the value we create: for accelerating your pull requests that merge to production, we'll bill around that. Per seat for now. And with the advent of really smart open-weight models, which are more faithful to our harness, our cost per token this year is going to go through the floor. I don't want to margin tokens. If we're creating a lot of value for you, we'll bill appropriately.
How did you get funded? Self-funded or venture?
Venture, and inbound from X on the buzz of Taskmaster. A lot of people in the Valley were using it, so a bunch of people wanted to jump on. Justin Kan, the founder of Twitch, was our first seed investor, and we've had a great working relationship. Tim Shea from 500 Startups was actually a user, reaching out to us on X saying, can you fix this bug? He was keen to come in as well. We haven't burned through that. We're sipping pretty lightly, and we use Hamster to build Hamster, so the productivity gains and leverage ratio our team gets are incredible. We can ship a lot of product on pretty light resources.
How big is your team?
Ten. Four in the US, four in New Zealand, a co-founder in Montreal, and one more in Bangkok.
How are you dealing with time zones and communication?
The time zone is fine, because the difference between San Francisco and New Zealand most of the year is three hours, plus or minus a day. Our Fridays are their Saturday, but we get a lot of time together. I have a team member in Florida and my co-founder in Montreal, so I wake up, talk with them, and get a lot of productive work done. The Kiwi team comes online, we do our standup, they keep pushing, and I'm online until 11 p.m. because so are they. It's a big day, but there's always someone working. Development and design are mostly in New Zealand.
When you look to the future, where do you see your product going?
The market is moving toward whatever the latest buzz on X is, usually what Andrej Karpathy tweeted recently: from the knowledge graph, which is beautifully built into Hamster and very good at taking in all your context and floating it up to any model, to the workflow graph, where we've had skills and routines and loops for a while. The workflow graph lets us be declarative about the work and set it off on long-running tasks. Hamster is focused on product engineering for now. You could see how it would generalize later, but one ICP for now. I see longer-running, higher-level goal setting with AI, with a permission model: go do this for me, and it asks, should I do that once, or every time? Then it chains those workflows together. More useful, higher-leverage tasks that can be executed and automated for you. A lot of the team collaboration, pulling someone in on a brief, Hamster does for us automatically now, and we're dialing it to the point where it's useful and not annoying. So I see that extra level of autonomy, but permission-based, or built on trust to the point where you say, yeah, that works every time.
Do you see your team needing to get bigger, or are you going to be a ten-person billion-dollar company?
We definitely want the leverage ratio, and we've stayed small intentionally. Our appetite is bigger than our stomach, so we could be pushing faster. One thing we've had to face is that the team members are individually excellent and the team works together really well, but there's still a team bottleneck you can't get past. It's not 10x times 10x times 10x. You get a better outcome from the collaboration between a systems thinker and an empathetic, creative person, but there's not an unlimited leverage ratio. So the team will grow. I've managed a lot of engineering teams over the years, and the tiger team, a small composable team of maybe eight engineers, is such a nice unit. I'd want to add another one of those units, working on different components of the software or the business at the same time, rather than stacking individuals on. We did scale up a bit more, and in an age of AI we had to think: we're more like a football team, we need high performance and trust. We were starting to lose that composable unit, and we decided not to keep going to sixteen. Keep it tight, keep it delivering. Within weeks of that decision, communication was up and team trust was up. Follow-the-sun can be a blessing or a curse: if someone finishes their day, do they push that PR and close it off, or say, I'll ship that tomorrow when they're online? Small team wins. We call it the leverage ratio. I talk to the team about spinning a number of plates at a time, and the cognitive load of context switching. There's a maximum to keep craft and taste and deliver with consistency. Spin fifteen plates and you'll drop some. We learned those lessons ourselves with the AI leverage and landed on that number.
One thing we see is pods of seven or eight shrinking down to two or three, highly focused, very tight small teams.
There's the tiger team, more like four: traditionally business, design, engineering, and product. That's still true. Design thinking, empathy, creativity, how might we, jobs to be done, all of that is still true. In our language, we're all builders now, and you have systems thinkers and more right-brained people. If you can have respect and trust and work on that together as a unit, it gives you a better outcome every time. That said, over the weekend I built a new open source project by myself, headphones on, finished by Monday morning. An individual can peel out and do that. But a team still works together really well.
Every time I come to San Francisco, this city feels so far in the future compared to everyone else. Austin is future-adjacent, and the further away you get, the more in reality people are. With teams in New Zealand, Montreal, Bangkok, and SF, is it harder to keep everybody at the same cadence as the folks in San Francisco?
We have an internal Slack channel, hashtag discovery, and my co-founder and I probably spend too much time on X drinking from the fire hose that is the Bay-based industry. But we diffuse it. Austin sounds like the dreamy middle ground. It's a lot, all the time, but I wouldn't rather be doing anything else. It's like an IV. It's good to be grounded in realism and what people are actually looking for, even in software and product engineering. I assume it's my role to drink from the fire hose, diffuse the information, read the tea leaves, and figure out where things are going. My co-founder in particular is trying to figure out where things are heading on a three-month timeline, as opposed to what Karpathy tweeted today.
I was just talking to someone about how I'm having to shrink my future projections. Instead of quarters or years, I'm thinking in months and days.
In the goals section of Hamster we have what we call strategy on a page, and we use it ourselves. We have a direction meeting once a week, and you iterate your direction every week. Twelve-month timelines and quarterly product increments have gone away. You need to be able to say, we have live measures coming in on the strategy page, how are we doing against activation targets, what can we bump today? Then a discovery meeting once a week, where we take those hunches, and everyone on the team can come up with an idea, iterate it, do a prototype, so creativity comes from everywhere. Then you set delivery intention: as much as is going on, we're trying to hit this goal this week, and on a two-to-three week timeline here's where we're heading, but it all drives toward that direction.
How do you think about goal setting in today's AI world? Historically you set annual goals, quarterly goals, weekly goals, all tied to a long-term vision.
It depends on the org. Not everyone is a product engineering org. In the real world, software engineering supports a wider business with a cadence it needs to be true to. But within software engineering, you can be driven by data that informs your gut, not data only. You should be able to make a gut decision on the next bet on maybe a six-week timeline, and work with the rest of the business on the longer goals. That's the right amount of time to think ahead, and it breaks down into increments. The tiger team will tell you, don't put arbitrary dates on things, don't give me priorities, just let me cook. There's a middle ground: as a team, how are we deciding what matters most and what we're doing next? In an age of AI, you can ship something meaningful in a week or two with a team, something quite powerful. That's the right timeline for SaaS software. It can be very different for hardware. A team can stay focused and follow its nose pretty well now.
One thing I'm seeing teams struggle with is putting quarterly or annual objectives up for the board meeting, and then something monumental happens in AI. It's not a hard right turn, but it's iterative, and at the end you say, the reality is everything changed.
The good news is the artifact you present to the next board meeting is so easy to write. If something monumental happens, you say, Claude, update my 56-page business plan and take this tweet into account, and it does. It probably becomes more performative. You have to get to real talk. In my previous life, I ran a software consultancy for ten years. We'd place 16 or 20 people in a bank or another regulated industry, and my job was talking at board level, explaining to a 3,000-person enterprise how 20 people can make a material difference and operate differently. It takes time. You have to do it incrementally. The consumerization of enterprise technology is always interesting: board members all have their iPhones and expect everything in the company to run like their iPhone, so there's a friction or inertia problem with how fast the enterprise can work. People want that, but they want good controls, risk management controls. To the extent you can build a framework that allows gut decisions but learns fast, that's what an enterprise may struggle with at different levels of politics or management. Often it's the board saying, we want this tomorrow, why can't we move this fast? It's not the board that has a problem with vision or hustle. It's the second and third tier interpretation and organizational structure that slows it down.
What's been the most fun aspect of your job with this company?
I get a kick out of delivering software that people use, where they say, yes, this makes my life so much easier. When we do releases that are a step change up, that gets me going. It gets me up in the morning. So I'm in the right space.
How do you get that feedback? Usage data, or people emailing you?
A community Slack with customers gives immediate feedback, and my team lives in there too. If a customer says, it would be great if the software did this, we can ship it that day or the next day. They come back and say, oh my God, that changes everything. That's cool feedback to get, and the cycle time is incredible. I was software consulting with enterprises for a while. You could do an 18-month project in a regulated industry, and the people doing the work didn't get any feedback or get the customer's hands on it for 18 months. It didn't matter when you left for the day in month three. When you change that cycle time and you're shipping to production multiple times a day, it's a much better flywheel.
Where do you see companies messing up? You have a lot of product organization insight.
Where I've seen companies mess up the most, across all sizes, is not giving teams permission, autonomy, and agency to be self-delivering. Any politics getting in the way, any ego getting in the way, the usual human team stuff. You've really got to build the trust that the person next to you is going to deliver, and that they'll make a mistake, shoulder it, learn together, and move on. Anything that's not that, and this is hard to say in a bigger organization, is shooting itself in the foot. Particularly with the leverage of AI, if you ask, why aren't we getting that leverage? Collaboration, trust, autonomy, agency, they're the key pillars. I could talk for hours about the hundreds of ways people don't do that. You also need to intentionally set direction as a team, so there's buy-in. The old business person or project manager would come in with a Gantt chart and arbitrary dates no one was bought into, and the team just kept chipping away. Not doing an intentional discovery process up front, figuring out as a team what to build, bringing a proof of concept, a whiteboard, a branch, and reasoning on it together until we all agree and understand what we're going to build. There was a time for planning poker in a standup or a backlog grooming session, but we're no longer at that level. We have tasks and subtasks, but that's what the AI does. Everyone's an engineering manager now. You have agency and autonomy to deliver that how you want, but the initiative you're driving toward is where you collaborate and figure out what to build. People mess up by not having those new ceremonies and working in a new way, and by not giving everyone the agency to do what they do best.
Normally in these interviews people hear, I built this open source project, 28,000 stars, Justin Kan reached out, everybody's giving me money. They see the gloss. What are some hard things you've gone through? Any bad misses?
Let me be the first to tear that apart. Our investors are great and they point us in the right direction. They rightly ask, how many customers are you listening to, how's traction? The right metrics. We went out in July, started building in August, and by October we had something suitably embarrassing to put in market. We sent it to a list of people, and the response was, meh. A lot of support from a great community, but meh. In software engineering now, people are very picky about the software they use. If it's not perfect, if you put 15 minutes in and get five minutes of productivity back, you never look at it again. Whereas if you put 15 minutes in, upload some stuff to the knowledge graph, connect some things, and it just knows you, now you're getting leverage. The biggest response we get from PMs when we pitch the idea is, I don't have time. Plenty of examples in my career of product managers saying, yes, I understand that if I do that little thing now it'll make our lives easier, but I don't have time, we've got to deliver this thing right now. Every PM feels that. Every PM has seen their engineers get a leverage ratio and go faster. I don't think PMs want a CLI, but they're using Claude Code because that's where they can put the skills. I don't think that's the preferred way of being.
What do you think is the preferred way?
Collaboration, back in more business-like tools. Using Anthropic as an example, Cowork and Claude Code are completely separate products that don't share much. You've got your engineering tool and your business tool, and PMs are trying to sit between the two. But to come back to it: when you build open source software, you get the buzz. There's a lot of hard work, long days, the SF lifestyle, pretty much 996. I don't expect anyone else on the team to do that, but that's my job. Then there's launching something to people, going ta-da, and everyone saying, okay, whatever. That's hard. But that's product. The job is to launch something. I think Michael at Cursor launched Cursor four times. There's a history of posts on Hacker News of launching an email program, and then it turned into Cursor and the rest is history. That's the game. You've got to be prepared for people to say, this sucks. Then: how does it suck? Tell me more. I do think we've built a platform that's really cool, and our ability to iterate and solve process problems for people is powerful. Back when I was building websites for my teachers in high school, I figured websites were like rocks in a river and the visitors were the water. You build the rocks, then you figure out how the water flows, pay attention to it, and move with it. That's the process we're in now. It can be brutal, but it's the hardest thing you do that makes it rewarding.
Something profound for young founders to hear is perseverance through all of that. There will be setbacks, and people will trash your product and say it sucks. The perception on social media is that everybody just does things and it hits. They try something, it fails, and they think, oh, I suck. No, we all suck. You just haven't sucked enough.
It's hard. A VC I talked to recently said they like to back people who can chew glass and get through the pain. I've had a few of those seasons in my life. You've got to have grit and determination. I think Elon said you need an exceptionally high pain tolerance. That's true, and you need the surroundings and support to get through it. It helps that I enjoy chewing glass. But the upside at the end is potentially massive scale and making a real impact, which to me is the most important thing. Not building software no one uses, but building something that makes a meaningful difference to a lot of people.
You're in the heart of it in San Francisco. Token costs are going down, and there's a hardware and chip sell-off happening right now. Who do you think are going to be the winners and the losers?
Massively expanding TAM, so I don't think it's winner takes all. There's an interesting crunch with open-weight models now against the hundreds of billions of dollars that have gone into frontier models. I think the harness is going to win: vertically specific harnesses. Claude as a model is nothing for engineers without Claude Code. They didn't just hand you a model and a CLI, they built a harness around it. Hardware vendors are going to make a lot of money, because the number of tokens used in the world is going to explode exponentially. I had a thesis this year that token speed would go up 10x and token costs would go down about 100x, and I think that's going to be about right by January. When Hamster had to use a frontier model at $25 per million tokens, it'll be 25 cents in January.
Is that because of the open source models?
Partly, and the inference and power costs. There are a lot of FPGA and ASIC vendors building chips for much faster inference. Training still needs a million Nvidia GPUs, but inference is massively more power efficient on the stuff coming online: SambaNova, Cerebras, all these people building very efficient inference chips, even chips built for a specific model. I was happy to pay $25 per million tokens six months ago, and I'll be very happy to pay 25 cents for something ten times smarter by January.
But how do the training models make up for the investment?
The spicy take this week is everyone except Anthropic signing the letter to the US government saying we're all for open models. You've got Jensen at Nvidia saying bring on the open models, they'll all run on Nvidia, and AMD releasing competing products. A lot of people are coming to the table, and it's a huge market. Some not-quite-public companies are going to get their valuations clipped. But Anthropic has gone from 10 billion to 75 billion in ARR in six months. They're doing okay. They'll have to keep proving value in the enterprises they're taking on, and they'll get a lot of that upside. So winners and losers: I don't know if anyone's going to lose. Traditional SaaS has already had its legs cut off. It's an ever-expanding market that's going to keep taking over other services revenue. Taking a broader lens on the impact of AI, I think the acceleration is going to be faster than what society is used to, and society, people, government, will have to respond. I'm a big fan of running a country like a family trust, not like a business. You've got to make sure the free market is working for everyone. There's going to be some aggressive removal of jobs people are used to doing, but I don't think it means everyone will be without a job. People can retrain. I don't think we'll be slaves to AI. AI is doing the toil, so we can be more creative, and people should train toward that as much as they can. I'm not a doom and gloom person. I'm an AI maximalist, because there's so much benefit, but not without empathetic change management.
I completely agree. I don't see it as the sentient overlord people are afraid of. It's a powerful tool. One thing I tell my team is, more is more. It's going to be different, but it's going to be more.
I don't think government has the answers to everything, but society has to look after those who are impacted negatively, at a much greater pace than it's used to. That's going to be scary. I watch a lot of podcasts with people who are big in the AI space, and they're humans too. I think our dear leaders building the AI understand that if they leave everyone behind, there will be pitchforks. That's serious. Does the opportunity diffuse equally? That's what everyone is asking. I genuinely believe there's so much opportunity, and I'm very excited by technology making humanity better. With great power comes great responsibility.
Couldn't agree more. Andrew, thank you so much. This has been such an interesting conversation.
Thank you. Good to be here.
Be part of the
conversation.
Whether you’re a CEO or founder who wants to be featured or a company interested in sponsoring, there’s a place for you here.