Who you're meeting
Group Partner at Y Combinator, where he has done more than 1,000 office hours with founders across four batches[1]. He sits on the founder-facing side of YC, not the investment committee, so this is a credibility-and-fit conversation rather than a cheque. He describes himself, drily, as a "recovering tech founder" and is based in San Francisco[4].
He has built two billion-dollar companies: GoCardless (Direct Debit payments, YC S11) and Monzo, the UK challenger bank, for which he was awarded an OBE in 2019[1]. He joined YC as a visiting partner in 2021 and went full-time in 2023[1].
On what he rates in founders: "perseverance and tenacity are incredibly important... we often see supergeniuses who give up too early"[6]. He responds to directness over polish, and he is a builder first, which shapes how he will hear a pitch.
- He has spoken very openly, in public, about mental-health struggles and even death threats during the Monzo years[5]. It is on the record and he chose to share it, but it is not small talk.
- His views on AI replacing jobs are deliberately provocative — he has said himself that a post was "optimized for engagement"[7]. Engaging him on the substance is good; treating the provocations as settled fact is a trap.
Their company
Y Combinator is the startup accelerator launched in 2005 that has funded more than 5,000 companies — Airbnb, Stripe, Coinbase, DoorDash, Dropbox and others — with an aggregate valuation approaching $1 trillion[8]. It runs a three-month programme four times a year and invests $500k per company[9].
For this conversation the important detail is that YC is not just funding AI companies, it is building with AI itself: "Over the past year, we have spent about half of our time working on AI agents... we now have fully deployed agents that automate customer support, manage our events, and help us make sense of our data"[10]. They practise what they fund, which means Tom will judge any claim about agents against what he has actually shipped.
Key people
Decision-path labels are inferred from public roles. YC runs a flat partner model, so there is no formal approval chain to map; confirm any warm intro before leaning on it. If a formal partnership structure ever came up, Jon Levy (MD, Partnerships) would be the relevant name[11].
What's recent and material
- May 2026 — Tom's most-viewed recent talk, "How to Build a Self-Improving Company with AI", coined "burn tokens, not headcount" and claimed YC companies now reach demo day with "about 5x more revenue per employee than 18 months ago... the constraint is token consumption"[2].
- May 2026 — Multiple W26 companies were funded specifically to fix agents that break in production — BentoLabs, Salus, dari.dev ("most agents break mid-task... cannot survive a 30-minute task")[3].
- Summer 2026 RFS — YC's thesis: "AI has stopped being a feature and started being the foundation", with a whole category for "Software for Agents"[15].
- May 2026 — OpenAI offered "$2 million worth of OpenAI tokens to every startup in the current class"[16].
- 2026 — The accelerator field is crowding fast: a16z Speedrun, South Park Commons' $500M "anti-accelerator" fund, Station F's OpenAI/Anthropic-backed F/ai[17], while AI compresses startup timelines so far that YC is being described as evolving "from an accelerator to a full operating system"[18].
Two signals make this the moment. First, AI is no longer a differentiator in the portfolio — "92% of companies in this batch incorporate AI", so the moat has moved to whether agents actually work in production[19]. Second, Tom has said the quiet part out loud: the self-improving company he advocates is still aspirational — "I am not sure anyone has a truly self-improving company in every function"[20]. The gap between the vision he is selling founders and the reliability reality is wide open, and that gap is precisely what Serpin closes.
Your angle into this meeting
The skill's read of the facts above — interpretation, not new facts.
- Likely problem — the companies Tom advises hit the wall where an agent demos well but fails in production, exactly the niche a wave of his own W26 portfolio is racing to fill[3], and he concedes even the self-improving model is not yet real anywhere[20].
- Likely cost — founders burning their scarcest resource, time, on reliability firefighting instead of product, when AI has made the differentiator product velocity not the AI itself[19].
- What good looks like — a repeatable way to take agents from "works in the demo" to "robust, reliable, responsible in production", which is the sentence Serpin already leads with.
- Thematically he is the perfect match: his "self-improving company" and "burn tokens not headcount" framing is the same world as Serpin's Agent Discovery and Design and Bounded Agency[2].
- As a buyer he is not a clean fit: YC is a network, not an operating company with the friction Serpin fixes, and Tom advises founders rather than procuring services. Treat the value as reach and credibility, not a contract.
- Bounded Agency speaks directly to the agent-reliability wall his portfolio keeps hitting[3] — governance and guardrails are how "works in a demo" becomes "safe in production".
- Serpin's practitioner stance (real operating experience, not just advice) lands with a builder who rates tenacity and dismisses armchair theorists[6].
Questions to ask
- You have said no one has a truly self-improving company in every function yet[20] — where do you see founders getting stuck when they try to take an agent from a working demo to something they can actually trust in production?
- With 92% of the batch building on AI, the AI itself is no longer the differentiator[19]. What separates the companies that pull ahead now?
- A real slice of W26 is companies founded just to stop agents breaking mid-task[3]. Is that a tooling gap that closes on its own, or a discipline founders have to build in deliberately?
- When you tell founders to "burn tokens, not headcount"[2], what is the failure mode you see when they take it too literally?
Objections to anticipate
- "Founders can just figure this out themselves." — Likely, given his faith in small high-agency teams[2]. Counter with the time cost of learning reliability the hard way, not a capability gap.
- "This is a solved problem, the tooling will catch up." — His own portfolio suggests it is not solved yet[3]; agree the tooling is improving, then separate tooling from the design discipline around it.
- "We do not bring in outside advisors." — YC's culture is self-reliant and builder-led[1]. Position Serpin as a pattern founders can own, not a dependency.
Opener and next step
- Open on the gap he named himself: he tells founders to build the self-improving company, yet says no one has one that actually self-improves in every function[20]. That gap is the whole conversation.
- Open on the portfolio signal: a chunk of his own batch exists because agents break in production[3] — "the cohort went AI-native faster than the discipline to ship it reliably did."
Offer to write up the move from "agent works in the demo" to "agent is reliable and governed in production" as a short pattern his partners could hand founders. No commitment, just useful, and it earns a second conversation without asking for one.