Antioch Agent is the product name attached to the most interesting idea in robotics tooling this year: describe a test in plain language, and a machine builds the digital twin, writes the scenarios, runs them by the thousand, and reports what broke. Antioch, the New York company behind it, raised a $32 million Series A on 8 September 2026 to build exactly that. The idea is real, the customers are named, and the engineering is not vapour.
What is missing is the documentation. We fetched every page Antioch publishes — ten product and company pages, two legal pages, both blog posts, the press index and the machine-readable index the company wrote for AI assistants — and counted what they say. Across 5,698 words of product and company copy, the word “browser” appears zero times. The product name “Antioch Agent” appears eight times, on five pages, carrying five different job descriptions. The 1,521-word post announcing the $32 million round never names it at all.
The browser workspace does exist. It sits at console.antioch.com, it returns HTTP 200, and its page title is “Antioch Console”. You reach it from one footer link. This article is about the gap between those two facts, because that gap is what a robotics team hits when it tries to evaluate the tool. We are broadly positive on what Antioch is building — the autonomous AI agents pattern applied to physical hardware is genuinely overdue — but a buyer comparing simulation vendors has to work from published material, and here the published material and the shipped product describe different things.
So this piece counts. Every figure below came from a page we fetched on 9 September 2026, and every count is reproducible with the same fetch. Where Antioch states a number, we use Antioch’s number. Where the arithmetic is ours, we show it. For teams already running computer vision stacks and weighing whether to move validation into simulation, the counting is the useful part.
Table of contents
- What Antioch Agent Actually Is
- Where the Browser Claim Comes From, and Where It Does Not
- Counting Every Antioch Agent Mention Antioch Publishes
- The Five Jobs Antioch Agent Is Given
- Reading the Antioch Agent Workspace Mockup Closely
- The Only Numbers Antioch Publishes About Running Simulations
- Local Authoring Sits Above Antioch Agent on the Same Page
- What the $32 Million Series A Says About Antioch Agent
- How Loudly Each Antioch Document Talks About the Agent
- What Antioch Agent Means for Teams Evaluating Simulation
- What the Antioch Agent Record Does Not Yet Tell You
- Frequently Asked Questions About Antioch Agent
- References
What Antioch Agent Actually Is
The company underneath it
Antioch Inc. was founded in 2025 and is headquartered in New York. Five co-founders: Harry Mellsop, previously on Tesla’s Autopilot vision team; Alex Langshur; Michael Calvey; Colton Swingle, previously at Google DeepMind; and Collin Schlager, previously at Meta Reality Labs. Three of them — Mellsop, Langshur and Calvey — earlier founded Transpose, a security and intelligence platform acquired by Chainalysis in 2023. The Robot Report put headcount at eight in April 2026.
Where the name sits in the product
On the Antioch homepage, the platform is described in three numbered sections. Section 01 is “Author simulations locally, run them in the cloud.” Section 02 is “High-fidelity simulation modules off the shelf.” Section 03 is “Automate development with Antioch Agent.” The product name therefore belongs to one third of the stated workflow, and it is the third of the three.
What Antioch Agent is claimed to do
Section 03 carries exactly two bullets. The first: “Onboard your system from CAD, BIM, spec sheets, and more.” The second: “Use the Antioch agent to control the simulation, design scenarios, and iterate on your autonomy stack.” That is twenty-six words of product description on the homepage, and it is the densest description of Antioch Agent anywhere on the site.
The surrounding platform
Around that sits a conventional simulation business. Antioch calibrates digital twins to customer hardware, runs scenarios at cloud scale, and integrates with NVIDIA’s Omniverse libraries, Isaac Sim and Isaac Lab, with infrastructure from Nebius. Named users are Amazon’s Ring, quoted through Jason Mitura, VP of Software Development at Amazon and Ring’s Chief Product Officer, and Launchpad Build AI, quoted through CEO Jon Quick.
Where the Browser Claim Comes From, and Where It Does Not
The console is real
console.antioch.com resolves and returns HTTP 200. Its <title> is “Antioch Console”. The served document is a 3,850-byte shell: a React application built with Vite, deployed on Cloudflare Pages, with an empty <div id="root"> that the JavaScript fills in. It reads a localStorage key called antioch-theme before first paint. This is a browser application, unambiguously.
What the console loads tells you what it is for
The shell’s own HTML comments name its rendering stack. It pulls in Plotly 2.27.0 — the comment says “Plotly.js for interactive charts from Jupyter notebooks” — plus Vega 5, Vega-Lite 5, vega-embed 6, and MathJax 3 “for LaTeX rendering in Jupyter outputs.” A console that ships a Jupyter output renderer is an analysis surface: somewhere you read evaluation results, not only somewhere you launch them.
The word the site never uses
Against that, here is the count. Across the homepage, four solutions pages, the blog index, both blog posts, the press page and the careers page — 5,698 words in total — the string “browser” appears zero times. It appears seven times in the 2,649-word privacy policy, and all seven are cookie and Do-Not-Track boilerplate: how to set your browser to reject cookies, and a note that Antioch does not respond to browser “do not track” signals. None describes the product.
One footer link, and no prose
The console is linked from the site exactly once: a text link labelled “Console” in the footer, under a heading marked “Product”, carrying target="_blank". Beside it is a “Trust center” link to trust.antioch.com, next to AICPA SOC and ISO/IEC 27001:2022 badges. There is no screenshot captioned as the console, no tour, no sign-in call to action, and no sentence anywhere explaining that Antioch Agent is something you drive from a browser tab.
| Antioch page | Words | “Antioch Agent” | “browser” |
|---|---|---|---|
| Homepage | 295 | 2 | 0 |
| Solutions: aerial autonomy | 204 | 3 | 0 |
| Solutions: ground autonomy | 339 | 1 | 0 |
| Solutions: industrial autonomy | 393 | 1 | 0 |
| Solutions: intelligent perception | 359 | 1 | 0 |
| Blog: Series A announcement | 1,521 | 0 | 0 |
| Blog: Autoresearch deep dive | 1,989 | 0 | 0 |
| Blog index, press, careers | 598 | 0 | 0 |
| Product and company total | 5,698 | 8 | 0 |
| Privacy policy (cookies, DNT) | 2,649 | 0 | 7 |
Counting Every Antioch Agent Mention Antioch Publishes
The method
Antioch publishes an llms.txt at its root — a machine-readable index written for AI assistants. It lists every page on the site and states that each URL also returns Markdown when the request carries an Accept: text/markdown header, with the response size reported in an x-markdown-tokens header. We fetched all twelve listed pages plus one blog post linked from the Series A article but absent from the sitemap, took the Markdown, stripped the trailing source line, and counted with a script.
Eight mentions, five pages
The result: the exact phrase “Antioch Agent” or “Antioch agent” appears eight times. Two on the homepage, three on the aerial autonomy page, and one each on ground autonomy, industrial autonomy and intelligent perception. That is the complete public footprint of the product name. Every other page — the blog index, both blog posts, press, careers, privacy and terms — has none.
The two posts that explain it without naming it
This is the sharpest number in the count. Antioch has published two blog posts totalling 3,510 words, and both are substantially about the agent. Neither uses the product name once. The Series A post, 1,521 words, calls it “an engineer or agent.” The April deep dive, “Autoresearch for Physical Autonomy” by co-founder Alex Langshur, runs 1,989 words about what “the agent” does and never once brands it.
Why that matters for search
A robotics lead searching for the product by name reaches five pages, four of which are vertical landing pages with one sentence each. The two documents that would actually answer “what does this thing do” are invisible to that search. Antioch’s own llms.txt compounds it by telling assistants the site “has no API to call, nothing to buy and no account to sign in to” — a description that is true of the marketing site and false of the console linked in its footer.
The Five Jobs Antioch Agent Is Given
Onboarding, on the homepage
Homepage bullet one gives Antioch Agent an ingestion job: “Onboard your system from CAD, BIM, spec sheets, and more.” Homepage bullet two gives it a driving job: control the simulation, design scenarios, iterate on the autonomy stack. Two bullets, two different jobs, no overlap in wording.
Scenario construction, on the aerial page
The aerial autonomy page runs a section headed “The Antioch agent: Describe the scenario. Antioch Agent builds it.” The body: bring CAD, URDF, sensor spec sheets, a ROS container and a flight controller, describe your test cases, and “Antioch Agent builds a digital twin, environments, and scenarios.” This is the most concrete published description of the product, and it is on a vertical page rather than the product section.
Variant generation, on the ground page
Ground autonomy gives it a narrower job under the heading “breadth”: “Antioch Agent generates scenario variants across the dimensions you care about: site layout, lighting, sensor degradation, SKU mix, human traffic.” Here the twin already exists and the agent is a combinatorial expander, not an onboarder.
Configuration selection, on the industrial page
Industrial autonomy gives it a fourth job: “The Antioch agent: Select and tune the right autonomy configuration.” The body describes comparing hardware, perception stacks and AI policies across vendors, and assessing safety, reliability and throughput. That is a procurement and benchmarking role, distinct from building anything.
Prompt-to-scene, on the perception page
Intelligent perception gives it a fifth: “Antioch Agent builds thousands of realistic scene variants with a single prompt.” The phrase “a single prompt” is the closest the site comes to describing an interaction model, and it appears once, on the fourth vertical page, 359 words long.
| Page | Job the page assigns | Inputs named |
|---|---|---|
| Homepage, bullet 1 | Onboard the system | CAD, BIM, spec sheets |
| Homepage, bullet 2 | Control sim, design scenarios, iterate | None stated |
| Aerial autonomy | Build twin, environments and scenarios | CAD, URDF, spec sheets, ROS container, flight controller |
| Ground autonomy | Generate scenario variants | Layout, lighting, sensor degradation, SKU mix, traffic |
| Industrial autonomy | Select and tune autonomy configuration | Hardware, perception stacks, AI policies |
| Intelligent perception | Build scene variants | “A single prompt” |
Reading the Antioch Agent Workspace Mockup Closely
Six steps, one instruction
The homepage’s section 03 illustration is the only place the interaction model is shown. Its frame is labelled “Agent workspace”, the project is “warehouse-amr”, the status is “complete”, and the conversation is marked “6 steps”. The typed instruction reads: “Onboard our warehouse AMR from cad/amr_v4.step, then design and run the docking regression suite.” Below the transcript sits an input placeholder: “Ask the agent to change anything…”
What the transcript claims
The steps are specific. “Thought for 6s”, then Read(cad/amr_v4.step), then an import from STEP to a URDF twin resolving to 4 links, 2 drive joints and identified masses. Then two writes: scenes/warehouse_main.usd and scenarios/dock_approach.py, producing 12 scenarios across approach, align, dock and retry. Then a run of 512 replicas at 96% passing — 491 of 512 — with failures clustering at fork offset of +2 cm or more. Then a fork-offset sweep of 512 replicas returning 512 of 512.
The detail that settles the browser question
Alongside the transcript the mockup shows a “Viewport” panel displaying warehouse_main.usd. A conversation pane, a run log and a 3D viewport side by side is a web application layout, and the aria-label Antioch wrote for the image confirms the intent: “The Antioch agent onboarding a warehouse robot from CAD in a conversation as the simulation viewport assembles alongside, ending with a 512-replica docking suite passing at 100 percent.” The evidence for a browser product is an image description, not a sentence of copy.
What a mockup is and is not
None of this is a benchmark. The 491-of-512 figure, the +2 cm fork offset and the 12 scenarios are illustrative values inside a marketing graphic, and should be read as a description of the intended workflow rather than a published result. That distinction matters, because these numbers and the section 01 numbers below are the only quantities Antioch attaches to Antioch Agent anywhere.
The Only Numbers Antioch Publishes About Running Simulations
The suite in section 01
Section 01’s illustration shows a scenario file, crosswind_approach.py, decorated @scenario(replicas=2048), dispatched to a suite named “shipyard-ops”. The dashboard beside it reports 2,048 replicas, an estimated cost of $4.10, and a pass rate of 95%, across eight named scenarios of 256 runs each.
The arithmetic checks out
Eight scenarios at 256 runs each is 2,048, which matches the stated replica count. The eight published pass rates are 97%, 93%, 98%, 88%, 96%, 99%, 91% and 100%. They sum to 762, and 762 divided by 8 is 95.25% — which rounds to the 95% the dashboard displays. Antioch’s illustration is internally consistent, which is more than most vendor graphics manage.
Here are those eight scenario pass rates as published on the Antioch homepage.
Two-tenths of a cent per replica
The $4.10 estimate against 2,048 replicas works out at $0.002 per replica — two-tenths of a cent for one run of one scenario. It is the only price of any kind on the Antioch site: there is no pricing page, no tier list, and no rate card. Whether that figure is representative of a customer’s real bill is not stated, and a crosswind approach is a cheaper thing to simulate than contact-rich manipulation.
What is not costed
Antioch Agent itself carries no price signal at all. Nothing on the site says whether agent-driven onboarding is included, metered separately, or bundled into an enterprise agreement, and the only route to a number is the “Book a call” button. For teams building a business case, that is the missing input.
Local Authoring Sits Above Antioch Agent on the Same Page
Section 01 says “locally”
The tension inside Antioch’s own homepage is worth naming. Section 01 is headed “Author simulations locally, run them in the cloud,” and its first bullet reads “Write Isaac Sim / Lab code from your local development environment, no GPU required.” Its image description says a scenario is “being authored locally in a code editor, dispatched to the cloud.” That is a local-first workflow, described first and in more detail than Antioch Agent, which appears two sections later.
The deep dive agrees with section 01
The April autoresearch post reinforces the code-first reading. It describes an interface “that an agent can drive as directly as a coding agent drives a Python interpreter,” says the platform hands back “a Docker image,” says every scenario becomes “a versioned artifact in your repo,” and says the suite “fires on every commit.” Repositories, commits, containers and CI are DevOps primitives, not browser ones.
Two workflows, one product, no map
So Antioch describes two ways in: write Python locally against Isaac Sim and Isaac Lab, or talk to Antioch Agent in a workspace. Both are plausible and they are not mutually exclusive — most modern developer platforms offer both. What the site never publishes is how they relate. Does the agent write into the same repo you author in? Does the console show runs you dispatched from your terminal? Nothing answers that.
Why a buyer needs the answer
This is not a pedantic point. If Antioch Agent operates on your repository, it slots into existing review and CI practice and your simulation scenarios are versioned alongside your autonomy code. If it operates only inside a hosted workspace, scenario definitions live in a vendor system and export becomes a procurement question. Those are materially different products, and the published record does not distinguish them.
What the $32 Million Series A Says About Antioch Agent
The round
Antioch announced $32 million on 8 September 2026, led by Greylock, with A*, Category Ventures, Box Group and Icehouse Ventures participating. Greylock general partner Saam Motamedi joins the board. Named angels include Shyam Sankar, CTO at Palantir; Adrian Macneil, CEO of Foxglove; and Ian Andrews, an NVIDIA executive and former Groq executive. Sankar also backed the 2025 pre-seed.
The framing is a verifier, not an assistant
The announcement’s core claim is that Antioch is “building the verifier for physical AI: a development environment that predicts whether a proposed change will improve a physical system before it reaches hardware.” It leans on a $50 trillion figure for the physical economy and on Ring’s testimony that Antioch’s simulations “closely matched our physical test results, including in scenarios we deliberately held out of calibration” — a genuinely strong claim, because held-out calibration is the honest way to test a simulator.
And it never names the product
In 1,521 words the announcement uses the word “agent” three times, all generically: “an engineer — or increasingly, an AI agent”, and “an engineer or agent” twice more. The phrase “Antioch Agent” does not appear. A company that has just raised $32 million and has a named agentic product on its homepage announced the round without mentioning it once.
The funding arithmetic does not reconcile
The post states: “Together with the $8.5 million seed round announced earlier this year, this brings our total funding raised to $40.5 million.” That sum is $8.5m plus $32m. But Antioch also announced a $4.25 million pre-seed on 8 December 2025, reported by SiliconANGLE and The Robot Report. Adding all three announced rounds gives $44.75 million, $4.25 million more than the stated total. Either the pre-seed is folded inside the $8.5 million seed figure or it is excluded from “total funding raised”; the company does not say which.
| Round | Announced | Amount | Lead or named backer |
|---|---|---|---|
| Pre-seed | 8 Dec 2025 | $4.25m | Shyam Sankar (Palantir CTO) among investors |
| Seed | 16 Apr 2026 | $8.5m | Not named in company post |
| Series A | 8 Sep 2026 | $32m | Greylock (Saam Motamedi joins board) |
| Sum of announced rounds | — | $44.75m | Our arithmetic |
| Total stated by Antioch | 8 Sep 2026 | $40.5m | “$8.5m seed … brings our total to $40.5m” |
| Unreconciled difference | — | $4.25m | Exactly the pre-seed |
How Loudly Each Antioch Document Talks About the Agent
The chart below counts product-name mentions against page length, and the pattern is the inverse of what you would expect: the longest documents name it least.
The two longest documents name it zero times
The 1,989-word autoresearch post is the best writing Antioch has published and the closest thing to product documentation for the agent. It describes onboarding from CAD and URDFs, hardware trade-off sweeps, writing autonomy software, regression hunting and model evaluation — five capability areas in detail. It brands none of them.
The shortest page names it most
The aerial autonomy page is 204 words, the shortest page in the set, and it carries three of the eight mentions. Density of the product name is therefore highest where the explanation is thinnest, which is a reasonable description of the whole documentation problem.
What Antioch Agent Means for Teams Evaluating Simulation
The category question is settled; the product question is not
Simulation-first validation is not speculative any more. Ring’s held-out-calibration statement is a serious endorsement from a serious hardware organisation, and NVIDIA and Nebius partnerships are real infrastructure. If you build drones, AMRs, industrial cells or fixed perception, the case for moving coverage into simulation stands on its own, independent of any vendor.
Five questions the published record cannot answer
Before a pilot, we would want written answers to five things: does Antioch Agent write into our repository or a hosted workspace; how is agent usage priced against replica cost; what does the console let us do that the local path does not; what happens to scenario definitions if we leave; and what calibration evidence exists for our hardware class rather than Ring’s. None of these is answerable from antioch.com today.
Where the tool plausibly fits
For a team already writing Isaac Sim scenarios, the local path is the low-risk entry: your code, your CI, your review process, cloud compute for the parallel runs. Antioch Agent then becomes an accelerator on top of an existing workflow rather than the workflow itself. That framing matches Antioch’s own section ordering, even if the marketing emphasis runs the other way.
The signal worth watching
The most informative thing Antioch could publish is the same autoresearch post with the product name in it, plus one page describing the console. The company clearly can write — the blog is better than its category norm. Read our AI models and tools hub for how other vendors in this space document their agentic surfaces, because the gap here is editorial, not technical.
What the Antioch Agent Record Does Not Yet Tell You
No public trial, no documentation
There is no free tier, no sandbox, no docs subdomain and no API reference. docs.antioch.com, app.antioch.com and sim.antioch.com do not resolve; console.antioch.com does, and returns an application shell that requires whatever authentication its JavaScript enforces. Every published route to Antioch Agent goes through “Book a call”.
No independent test
No third party has published a hands-on review of Antioch Agent. The thirteen press items Antioch lists are funding stories and category pieces; none is a product test. The Ring quote is the strongest evidence available, and it is a customer statement in a company announcement, not an independent benchmark.
The honest summary
Antioch Agent is a real product with a real browser workspace, backed by $32 million and named customers, described by its maker in twenty-six words on a homepage and five one-line variations across four vertical pages. The engineering appears to be ahead of the documentation. That is a much better problem than the reverse, and it is still a problem for anyone trying to evaluate it this quarter.
Frequently Asked Questions About Antioch Agent
Is Antioch Agent actually browser-based?
In practice, yes. console.antioch.com returns HTTP 200 with the title “Antioch Console” and serves a React web application. But Antioch never writes the word “browser” on any product page, so the claim rests on the console’s existence and a workspace mockup rather than on published copy.
What does Antioch Agent do?
Depending on which page you read: onboards a system from CAD, BIM and spec sheets; builds digital twins, environments and scenarios; generates scenario variants; selects and tunes autonomy configurations; or builds scene variants from a single prompt. All five descriptions are Antioch’s own.
How much does it cost?
Antioch publishes no pricing. The only figure anywhere on the site is an illustrative “est. cost $4.10” for 2,048 simulation replicas in a homepage graphic, which is $0.002 per replica. Nothing states how Antioch Agent usage itself is priced.
Does it replace Isaac Sim?
No. Antioch integrates NVIDIA’s Isaac Sim, Isaac Lab and Omniverse libraries rather than replacing them, and the homepage explicitly describes writing “Isaac Sim / Lab code from your local development environment.”
How much has Antioch raised?
Antioch states $40.5 million total as of 8 September 2026. Its three announced rounds — $4.25m pre-seed, $8.5m seed and $32m Series A — sum to $44.75 million, so the pre-seed is either included in the seed figure or excluded from the stated total.
References
Antioch — High-fidelity simulation for Physical AI, at cloud scale
Antioch raises $32M to bring software speed to physical AI development
Autoresearch for Physical Autonomy
Antioch llms.txt — agent-readable site index
Antioch Raises $32M to Bring Software Speed to Physical AI Development
Antioch raises funding to bring ‘software speed’ to robot development
Antioch prepares to accelerate simulated testing for autonomous robots after raising $8.5M
Antioch raises pre-seed funding to accelerate AI robotics testing
This simulation startup wants to be the Cursor for physical AI
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.