The last two days funded three agents that each do exactly one thing — and that is the story. Between September 30 and October 1, the market priced offensive-security agents, hardware-engineering agents, and the optical plumbing that feeds them, and it priced all three steeply. Here is what happened, and what each item changes for a founder.
1. Armadin: Mandiant's founder raises $255.5M for agents that attack you first#
Kevin Mandia — who founded Mandiant and sold it to Google for $5.4 billion in 2022 — raised $255.5 million for Armadin at a valuation north of $2.5 billion. Armadin runs continuously-online "agent swarms" that chain multiple vulnerabilities together to execute real-world attacks, the way a human red team would — the pitch is to replace periodic, human-led penetration testing with an autonomous attacker that never stops probing. The round drew the names you would expect for a founder with that exit history.
The person who spent a career cleaning up breaches is now selling the breach — automated, continuous, and pointed at your own stack before someone else's agent gets there.
What it means: Offensive security just became an agent category with a nine-figure war chest behind it, and the framing matters more than the dollars. "An attacker that runs all the time" is a different product from "a scanner you run quarterly," and it resets the baseline customers will expect. If you sell anything security-adjacent, the question in every deal is about to shift from what do you detect to what would a tireless attacker-agent find first — and if you build anything that touches customer data, assume the tools probing you are now autonomous and continuous. The cheap move this quarter is to point one of these swarms at your own surface before a hostile one does; the expensive move is to find out the hard way that quarterly pentests were never the real cadence.
2. Flow Engineering: $50M to give hardware its own agent harness#
Flow Engineering raised $50 million at a $750 million valuation, with Valor Equity Partners and Atreides Management co-leading and Sequoia (which led the Series A) following. Flow is a system of record where engineering teams and AI agents design hardware together — requirements, design, and verification held in one connected graph — so frontier models can work securely against live, sensitive hardware programs. Reporting around the raise says Rivian grew from 40 to roughly 1,500 users on the platform in about seven months.
What it means: The agent-harness pattern that reshaped software is crossing into atoms, and the interesting part is why it needs its own tool. Hardware can't be vibe-coded against a repo; it carries requirements, tolerances, and verification that a general coding agent has no structured place to put. Flow's bet — and the bet a $750M mark endorses — is that every high-stakes engineering domain will want a harness that models the domain's own data, not a generic chat window bolted onto it. For founders, that's the reusable lesson: the defensible agent products right now aren't the models, they're the domain-shaped surfaces that make a model trustworthy inside one messy, consequential workflow. If your industry still runs on spreadsheets and tribal knowledge, that gap is the product.
3. Volantis: $88M to move the bottleneck from copper to light#
Volantis raised an $88 million Series A, co-led by Lachy Groom and Abstract Ventures, with John Doerr among the participants. Its technology replaces copper wires with laser beams to connect compute chips to memory chips: an optical fabric that pools memory and aggregates its bandwidth as you add more, wiring a single GPU to as many as 220 memory chips versus roughly 8 in today's setups. First integrated inference engines are slated for 2027.
What it means: The constraint on running agents has quietly moved. For two years the scarce thing was FLOPs; increasingly it's memory bandwidth and capacity — the wall you hit when a long-running agent's context, KV cache, and weights all need to live close to the compute. Volantis is a bet that the next performance unlock isn't a faster kernel but a wider pipe to memory, which is the same story the inference-optimization world has been telling from the software side. Founders don't need to buy optical interconnect, but you should read the signal: the economics of the agents you ship in 2027 will be set as much by memory plumbing as by model choice, and the vendors quoting you inference prices are about to have a new variable in their cost stack. Watch where the memory wall moves — it's upstream of your margins.
The thread#
Three raises, one direction. Nobody in this batch funded a general-purpose "do-anything" agent. They funded an agent that does one high-consequence job — break in (Armadin), build hardware (Flow) — and the wiring that decides whether such agents can run economically at all (Volantis). For a team of one, the instruction is sharper than last week's: generality is no longer the pitch. Pick a workflow where an agent doing one thing reliably has real stakes attached — money, safety, physical output — build the domain-shaped surface that makes a model trustworthy there, and price your future on the memory curve, not just the model curve. The capital has stopped paying for agents that can do anything. It's paying for agents that do the one thing you can't afford to get wrong.
