Introducing Emergent: The… | Newcomer Newsletter Summary

1. Key Themes

Theme: Agent infrastructure is the biggest category: tooling to run, secure, observe, and pay for agents

The newsletter flags this directly, and the nominations show the stack filling in layer by layer: payments, security, observability, reliability, and trust.

Agent tooling is the dominant category

"The biggest group of startups is building the tools to run and control agents. But hardware is also hot."
Felicis's own read: it "highlights agent-infrastructure startups and the crucial challenge of getting workflows right."

Payments and identity for agents

Agentcard: "Agentcard believes agents should have full autonomy to pay for things online." (Guillermo Rauch, Vercel)

Security and governance of agents

Arrakis: "Security tooling was built for humans using machines; agents now act autonomously inside CRMs and financial systems — and Arrakis solves it." (Mati Staniszewski, ElevenLabs)
Mnemom: "As agents do more of the world's work, trust in them will have to be earned from a record, the way it is with people, and Mnemom is built around that from day one." (Reid Hoffman)

Observability, evals, and reliability

Armature: "The new normal: users reach products through Claude, ChatGPT, and coding agents. Armature captures every session on your MCP server, Claude Connector, or ChatGPT App..." (Rauch)
Convergent: "optimizes agent deployments by mining failures from production traces to create high-fidelity evaluation environments, and deploying coding agents to directly fix observed issues." (Ryan Daniels, Crosby)
Ampersand: builds "a living model of its operations, which we use to rearchitect the business, eval agents against reality, and build automation."

Routing and developer-throughput bottlenecks

Martian: "As the number of models keeps growing, picking the right one for each task becomes core infrastructure, and Martian is ahead on it." (Brendan Foody, Mercor)
Peak: "AI agents now write code faster than teams can test it, and CI has become a critical bottleneck to throughput." (Gabriel Stengel, Rogo)

Theme: Open-source and on-device models as a cost and privacy lever

Several nominees build on open-weight models or run inference locally, framing it as a structural cost and privacy advantage.

Open-weight models as the cost moat

Newcomer: "Commitment to open source comes up again and again, either as a key way to keep costs down or for running queries locally on smaller devices like phones and laptops."
Hercules: "By using open-weight models to power their app builder, Hercules ensures customers can build sophisticated applications at a much lower cost." (Lin Qiao, Fireworks AI)
Phylo: "By customizing open-weight models for biology use cases, they give scientists the control and expertise they need at a fraction of the cost." (Qiao)

On-device and local inference

Mirai Labs: "runs text, voice, and vision models directly on phones and laptops - no cloud, no per-query cost, no latency."
Ghost: "running AI models locally to keep personal data private while continuously working in the background."

Theme: AI-native services replacing services businesses, not adding tools to them

The article calls this a major investor theme, and several nominees pursue it explicitly.

Services as the target

Newcomer: "AI-native services and support companies have been a big theme with investors this year and we're seeing plenty of them here."
Nowadays: "Event planning is mostly email and chasing, exactly the work agents are already good at, so it replaces a services business rather than adding a tool to one." (Staniszewski)
Mistos: "AI data centers are running out of power, and every new plant is gated by years of engineering. Mistos does that engineering with AI." (Naveen Rao, Unconventional AI)
Wonder Studios: "The bet is that the winner here is a studio: AI-native creators plus Hollywood craft, owning the IP, at a fraction of the cost per minute."

Theme: AI is moving into the physical world and the hard sciences

Hardware, robotics, and bio are all represented, with the common thread being closing the loop between models and real-world data.

Hardware and robotics

Newcomer: "hardware is also hot."
Icarus Robotics: "building AI-powered robots to handle physical work in space... tackling a problem where human labour is incredibly expensive and constrained." (Eoin Hinchy, Tines)
Enigma: "builds AI systems that make any robot intelligent and intuitive to use, with software designed to work across different types of robots."

Bio and neuroscience

Instance Bio: "models can call an API with a long list of candidate binder sequences and targets, and Instance will run that m × n analysis all in a single massively parallel reaction, which is both a better representation of biological reality and vastly more scalable." (David Luan, Adept)
Alljoined: "EEG approaches to BCI have been limited in the past, but Alljoined has shown that massive scale can improve effectiveness, similar to LLMs and world models." (Cristóbal Valenzuela, Runway)
Felicis flags "healthcare, bio, and creator tools."

2. Contrarian Perspectives

Perspective: Don't build wet labs for bio AI; change the unit of experimentation

The consensus approach to closing the experimental loop for bio models is automated wet labs. David Luan argues this is incomplete:

"Instance has convinced me that the right way to build a physical RL environment for bio models is not (only) by building automated wet labs."
"You actually need to change the atomic unit of experimentation away from a simple pairwise 'does this protein bind to this target' towards measuring the full distribution of interactions between m protein binders and n targets."

Perspective: The AI payoff comes from expanding employees, not just automating them

Against the pure-automation narrative, Reid Hoffman backs SkillBench on the premise that returns accrue to firms that grow their people:

"The biggest AI returns will go to companies that also expand their employees' individual potential."
"SkillBench helps people automate more tasks while growing into more valuable work that draws on their judgment, discretion, and relationships."

Perspective: The winner in AI film is a studio, not a tool

Rather than selling software to creators, ElevenLabs' founder argues owning IP and production wins:

"The bet is that the winner here is a studio: AI-native creators plus Hollywood craft, owning the IP, at a fraction of the cost per minute."
Evidence cited: "the Lewis Capaldi video with DeepMind and Universal, a documentary with Campfire."

3. Companies Identified

Etched

Description: AI chip company founded by Robert Wachen.
Why mentioned: Case study of the summit spotting winners early.
Quote: "he's since gone on to raise nearly $2 billion in funding (and will be speaking onstage next week)."

Physical Intelligence

Description: Robotics AI company; Chelsea Finn presented.
Why mentioned: Early summit catch after emerging from stealth.
Quote: "We had Chelsea Finn from Physical Intelligence present in 2024 after the company came out of stealth."

Factory

Description: Company led by CEO Matan Grinberg.
Why mentioned: Past attendee now speaking onstage.
Quote: "Matan Grinberg, the CEO of Factory, is also a past Cerebral Valley attendee who will be on stage this year."

Exa, Higgsfield, Wonderful, Fluidstack

Description: AI companies whose founders attended early.
Why mentioned: Evidence of the summit's early-spotting record.
Quote: "The founders of Exa, Higgsfield, Wonderful, Fluidstack, and many more came to the summits while they were still at the beginning of their companies."

Felicis

Description: Venture firm and partner on the Emergent list.
Why mentioned: Co-produced the list and published its own analysis.
Quote: "We're pleased to be working with Felicis on the list."

Emergent cohort (29 startups)

Agentcard: "powers payments for agentic startups, consumer apps, B2B companies, and crypto teams run their agents' purchases."
Agrippa Industries: "building the future of US maritime logistics."
Alljoined: "non-invasive hardware that decodes what people see directly from brain activity."
Almanac Health: "grounding clinical AI in reliable medical evidence."
Ampersand: "Founding team is ex-Palantir, OpenAI, and Google, and led the largest engagement in Palantir's history."
Armature: "shows what users ask, what agents think, and how they use product."
Arrakis Security: "discovers the AI agents running inside an enterprise, profiles their behavior, and blocks or revokes what they do in real time."
Attention Engineering: "Recording the context or 'analyst exhaust' that comes from actions you take on your computer."
Cartwheel: "turns video, text, and motion libraries into 3D character animation that fits within existing studio pipelines."
Convergent: "The team has worked across all parts of the agent reliability problem at DeepMind, Meta, Mercor, Perplexity, Glean, and similar companies."
Conway: "builds a foundation model for predictions."
Enigma: "make any robot intelligent and intuitive to use."
Ghost: "Its first product, Core, is a screenless device that connects to your apps, files, and smart home devices."
Hercules: "Just by chatting with an agent, customers can build business critical applications, like a CRM or inventory management system."
Icarus Robotics: "AI-powered robots to handle physical work in space."
Instance Bio: "run that m × n analysis all in a single massively parallel reaction."
Martian: "a model router that sends each request to the model best suited to it in real time."
Mirai Labs: "Founders built Reface and Prisma, so they've shipped on-device AI to hundreds of millions before."
Mistos: "the AI-native engineering firm for power infrastructure."
Mnemom: "every action is signed as it happens, confirmed against what the company asked for, and kept on an independent record the company owns."
Nowadays: "plans corporate events for companies like Google, Amazon, and Notion."
Peak: "usage is growing more than 10x month over month."
Phylo: "makes it easy for scientists to use AI to accelerate their research."
Resin: "generates new stems that fit the track an artist is already working on."
Scuba: "Noah just raised a seed round from Accel and IVP."
SkillBench: "building the intelligence layer that helps companies scale high-quality AI automation by developing the human expertise it depends on."
Tako: "aggregated a ton of data across many different sources and cleaned it up so LLMs can actually read it in sensible ways."
Tomo: "growing like crazy among a different demographic than the typical agent startup — people outside of Silicon Valley."
Wonder Studios: "an AI-native film studio in London producing commercial and original work for brands, labels, and studios."

Nominator companies

Vercel (Guillermo Rauch), Snorkel AI (Alex Ratner), Runway (Cristóbal Valenzuela), Glean (Arvind Jain), Crosby (Ryan Daniels), ElevenLabs (Mati Staniszewski), Mercor (Brendan Foody), Cartesia (Karan Goel), Fireworks AI (Lin Qiao), Tines (Eoin Hinchy), Adept (David Luan), Unconventional AI (Naveen Rao), Rogo (Gabriel Stengel), Abridge (Shiv Rao), Imbue (Kanjun Qiu), Recursive Superintelligence (Josh Tobin): mentioned as the founder-led companies whose leaders made nominations, e.g. "We asked current and former founder speakers to nominate seed-stage (or earlier) companies on the cusp of greatness."

4. People Identified

Robert Wachen

Description: Founder of Etched.
Why mentioned: Early summit speaker who scaled dramatically.
Quote: "we hosted Etched founder Robert Wachen at our second-ever summit in 2023, and he's since gone on to raise nearly $2 billion in funding."

Chelsea Finn

Description: Physical Intelligence.
Why mentioned: Presented post-stealth in 2024.
Quote: "We had Chelsea Finn from Physical Intelligence present in 2024 after the company came out of stealth."

Matan Grinberg

Description: CEO of Factory.
Why mentioned: Past attendee, speaking this year.
Quote: "Matan Grinberg, the CEO of Factory, is also a past Cerebral Valley attendee who will be on stage this year."

Aydin Senkut

Description: Founder and managing partner, Felicis.
Why mentioned: Shared takeaways from the list.
Quote: "Aydin Senkut, Felicis' founder and managing partner, broke down his biggest takeaways from the list on the Newcomer podcast."

Noah Weiss

Description: Scuba founder, former Slack chief product officer.
Why mentioned: Praised for taste and pedigree.
Quote: "Noah was chief product officer at Slack, he has exceptional taste and judgement - what it will take to stand out in that crowded space." (Shiv Rao)

Veer Bhatia

Description: Founder of Peak.
Why mentioned: Fast execution with a small infra team.
Quote: "Veer Bhatia and his small team of infrastructure vets ship remarkably fast." (Gabriel Stengel)

Reid Hoffman

Description: Founder, LinkedIn; nominator.
Why mentioned: Backed Mnemom and SkillBench.
Quote: "Trust in them will have to be earned from a record, the way it is with people."

Other nominators

Guillermo Rauch (Vercel), Alex Ratner (Snorkel AI), Cristóbal Valenzuela (Runway), Arvind Jain (Glean), Ryan Daniels (Crosby), Mati Staniszewski (ElevenLabs), Brendan Foody (Mercor), Karan Goel (Cartesia), Lin Qiao (Fireworks AI), Eoin Hinchy (Tines), David Luan (Adept), Naveen Rao (Unconventional AI), Gabriel Stengel (Rogo), Shiv Rao (Abridge), Kanjun Qiu (Imbue), Josh Tobin (Recursive Superintelligence): peer founders whose picks signal where operators see promise. Representative quote from Kanjun Qiu: "I use it as an API, and it works really well."

5. Operating Insights

Build for the bottleneck your category just created

Peak's thesis is that AI-accelerated code creates a new constraint downstream. Operators can look for the second-order bottleneck created by a technology shift.

"AI agents now write code faster than teams can test it, and CI has become a critical bottleneck to throughput."
"Teams switch by changing one line, and their builds run several times faster."

Make context a product: failure data, traces, and operator actions

Several companies turn exhaust (production traces, user sessions, analyst actions) into the asset.

Convergent: "mining failures from production traces to create high-fidelity evaluation environments."
Attention Engineering: "Recording the context or 'analyst exhaust' that comes from actions you take on your computer, to provide richer context and/or training signal."

Pick a segment the typical agent startup ignores, and clean the data others won't

Distribution and data hygiene can be differentiators.

Tomo: "growing like crazy among a different demographic than the typical agent startup — people outside of Silicon Valley."
Tako: "aggregated a ton of data across many different sources and cleaned it up so LLMs can actually read it in sensible ways."

6. Overlooked Insights

Founders' own picks as a leading-indicator signal

The list's methodology leans on peer judgment when metrics are thin, which is a useful sourcing heuristic.

"Identifying the most interesting early-stage startups is as much art as science, requiring a feel for the technology, the people, and the market opportunity when there aren't a lot of numbers."
"Collectively they give a good window into how founders think about their peers."

Power is becoming the constraint on AI, creating services opportunities

Mistos signals that engineering capacity for power infrastructure, not just chips, gates data center buildout.

"AI data centers are running out of power, and every new plant is gated by years of engineering."