Andreessen Horowitz and Atlas Holdings launched Foundry Management, a program that embeds a small group of top technologists inside Atlas's industrial portfolio to build AI companies from the factory floor up. The Atlas portfolio represents $26B in annual revenue across steel, metals and mining, automotive supply chains, energy, and logistics. The pitch: the hardest part of industrial AI is not the technology but getting close enough to the work, so technologists start without a finished idea, build inside real operations, and can spin out as founder-CEO with a16z and Atlas backing when the work shows independent-business potential. Recruiting runs through talent@foundrymgt.com.
Why it matters
This is the American Dynamism thesis turned into a company-building machine. The signal for founders: the next wave of valuable AI companies may not start as software at all, but as products born inside steel plants, mines, and logistics networks where the data and the pain live. The edge goes to technologists willing to work on the factory floor rather than in a pitch deck, and to investors who can offer distribution into real industry, not just capital. Watch for copycat programs from other industrial PE firms; the template is now public.
Ayman Nadeem, founder of Nuanced (W24), published a detailed post-mortem on building an entire desktop coding app around 'plan mode' - and why it failed. His four lessons: models got good enough at exploring codebases that explicit instructions shrank; AI-generated specs added information without adding clarity; the linear chat-to-spec-to-approve-to-build workflow forced users to 'finish thinking' too early, since real thinking mixes planning and building; and splitting plan and build modes asks users a question the AI should answer from context. The emerging loop he sees: understand, act, inspect, clarify, adjust, repeat. The unsolved problem: keeping humans oriented while hundreds of agents change a system at once. The post drew heavy engagement and crossed over to Hacker News.
Why it matters
The dev-tools winners will not be the ones with the most modes and approvals, but the ones that keep the human in the loop with the least friction. Planning is becoming a background behavior of the model, not a product feature. For builders: watch how much of your 'agent UX' is scaffolding for a weaker model that no longer needs it, and design for the moment a hundred agents are editing at once, because that is where the real interface problem is.
Petrichor (F26), founded by CEO Wilhelm Hedenskog, is building a wet lab that frontier AI labs can use like cloud compute: labs send experiments, Petrichor runs them in parallel, and returns data from every step. The thesis: code and math models improved fast because they can check their own answers, but in biology the answer only comes out of an experiment, so experiment throughput is the ceiling on model quality. Models submit experiments directly through a Python library; customers own all resulting data; everything is human-run for safety until robots are proven safe. The company is asking for intros to research-data teams at frontier labs, empty Bay Area bench space, and used lab equipment.
Why it matters
This is the clearest bet yet that biology is the next training-data frontier, and that the bottleneck is physical: pipettes, not parameters. Whoever controls high-throughput experiment capacity could become the data supplier for every bio foundation model, which is a powerful position. Watch whether labs actually outsource wet-lab work or keep it in-house; the answer decides if this is a platform or a service business.
Stacktrace (F26), founded by ex-Twitter, Databricks, Pinterest, AWS, Twilio, and Harness engineers (the CEO was a VP of Engineering), launched an agent reliability platform after an agent bug burned $40,000 in AI costs in a single day and took a week to trace by hand. Setup takes two minutes via pip, npm, or an OpenAI LiteLLM proxy, and the platform streams full-lifecycle telemetry that catches silent drift, cost leaks, tool-loop bugs, and prompt failures in real time. The YC offer: free Team plan for three months, the $199/month base fee waived, and 10x usage - 2,000 session-hours across 100 agents per month.
Why it matters
Agent failures are moving from embarrassing demos to real money, and that is exactly when monitoring companies get built. The $40k-in-a-day story will repeat at hundreds of companies as agents touch production, which makes reliability infrastructure one of the most defensible picks-and-shovels bets in the agent stack. For operators: if your agents can spend money or change state, you need this layer before you scale them, not after.
Routant (F26), founded by Aditya Mehta, tackles the agent-to-customer access problem: builders connect their agent once to Routant's private network through a single MCP endpoint, and each customer registers its own apps and resources with precise control over which tools the agent can use, which actions need approval, and a full per-customer audit record of every request and decision. An outbound-only gateway reaches on-prem systems without opening inbound ports, and the agent's runtime stays in the provider's environment. Mehta, a former quantitative options trader at IMC, is applying the identity, authority, and audit discipline of automated trading systems to cross-company agent deployments.
Why it matters
Agents selling to enterprises keep hitting the same wall: nobody will hand a vendor's agent the keys to their systems. Whoever solves cross-company agent access with real audit trails unlocks the enterprise agent market. This is the shape of that solution - identity, least privilege, and logs borrowed from finance - and it suggests the winners in B2B agents will look more like security companies than chatbot companies.
Boat (F26) spent nine months building a Devin-like product and its own sandbox, then pivoted to full VM sandboxes for long-running agents and launched battleships.dev, a comparison of 366 'cloud infrastructure for agents' providers. Nearly half are YC companies; more than 120 YC startups now sell in this space. The rankings are built from real usage: thousands of agent sessions on Boat's own cloud VMs, with a browser in each VM scraping pricing pages, public benchmarks, and company directories, so every claim on the site is backed by a quote from the provider's own public page.
Why it matters
When 366 companies, including 120-plus from a single accelerator pipeline, are selling the same cloud-computer-for-agents product, compute for agents is already a commodity, not a moat. The winners here will be whoever has the cheapest reliable capacity and the best developer experience, not the most clever wrapper. For founders: think twice before starting another sandbox company; the comparison site now exists. Connection: taken together with Stacktrace (monitoring), Routant (access control), and Spacesheep (workspaces), this is the week the agent stack got its boring-but-necessary plumbing.
Spacesheep (F26), founded by Michael Makarov, turns any agent session into a shareable workspace: tell any agent (Claude Code, Codex, ChatGPT, Gemini, or Grok, via Telegram, Slack, or iMessage) to 'ship this to spacesheep.dev' through one MCP server, and the session output becomes an interactive space the team can read, comment on, and iterate on. It supports running code, live-streamed data (one demo streams per-core server load twice a second), games, and A/B voting. Makarov was at Anthropic, where his side project became Claude Code Artifacts; he also worked at Pinecone and Twitter and was a founding engineer at Sonar (acquired). Early validation: Stepan Parunashvili of Instant (S22) commented that it 'just pointed my agent and it worked.'
Why it matters
Agent output is still trapped in chat windows, which is a terrible place for a team to work. The company that owns the shared surface where agents and humans collaborate owns a lot of attention, and Artifacts showed the pattern works. The risk: this is a feature the big labs could ship, so the moat has to be in the multi-agent, multi-model layer they will not bother to build.
A Bookface essay by Keanu Clark (Nxtcure Labs, F26) asked whether AI agents could one day hold corporate personhood - agents as CEOs, agent founders who exit and angel invest. The most concrete signal came in the comments: Mitch Duncombe of Vector Legal (W26) reports that Delaware lawmakers are considering legislation, possibly as soon as January, to create a new 'Artificial Intelligence Company' entity - 'like an LLC, but for agents' - with a 30-month regulatory sandbox and a permit application for testing and observability. Treat the legislation claim as reported, not confirmed; it comes from a commenter, not an official source.
Why it matters
Even as a rumor, this tells you where the smart legal minds think agent commerce is heading: toward entities that let software sign contracts, hold assets, and be audited. If Delaware moves first, it will pull agent businesses into its jurisdiction the way it did with LLCs. Founders building agent marketplaces should watch this closely, because the legal wrapper may matter as much as the model.
