The Cost of “It Works”

Hone

Hone wants a correction made in Slack to survive the conversation. The next person who works with its AI should benefit from something a colleague already taught it, without having to teach the same lesson again.

That is an unusually practical place to put ambition. A company spends years learning how to work, then watches that knowledge scatter across people, messages and decisions. Hone is building Engines that carry the learning into the next assignment.

The San Francisco team announced a $60M Seed on October 8, led by Benchmark and Index Ventures. Hone told Bloomberg it is valued at $285M. Elad Gil, Hanabi, Definition, Diffusion, Lux, SV Angel, and Align also participated, the company says.

Co-founders Moritz Stephan, Oliver Brady and Carlo Kobe bring different pieces of that operating problem. Moritz Stephan helped build Devin's planning and execution system at Cognition before becoming its chief of staff. Oliver Brady worked at Harvey and helped build Mercor's enterprise effort. Carlo Kobe co-founded Fizz with Stephan and kept building it while Stephan finished Stanford.

Their Engines are designed to take on continuing goals such as customer retention or vendor management. The organization sets success and boundaries; the software learns the relevant history, works with people, and keeps pursuing the outcome. A useful colleague still needs to know when somebody else gets to make the call.

Hone describes approvals for sensitive actions, action logs and routines kept in a versioned repository. Before an Engine changes its own behavior, the team gives it a way to replay past work. Learning from one mistake has to coexist with remembering how to handle the situations it already understood.

In 125 scenarios, Hone reports that replaying changes before promotion reduced regressions from 33% to 5%. That is an internal engineering result with a defined scope. Customer revenue, deployment costs and independently measured financial returns remain undisclosed in the launch materials.

The $60M finances building Engines, with hiring across research, infrastructure, product and deployments. There is real work in that list: connecting an organization, making its history usable, and deciding how a system earns permission to act again.

For a team that tries Hone, the relationship keeps developing with each correction. Someone teaches the Engine why a promise to a customer matters, and the next colleague inherits software expected to remember it.

Trmeric

Fifteen minutes can buy approval for an AI budget. The next year has to explain which part created value.

That gap is where Trmeric is building. The company raised a $6.5M Seed round led by Hitachi Ventures. Schema Ventures, CerraCap Impact Venture Capital, FalconX Ventures, Executive Venture Fund and three angel investors participated.

Siddharth Bohra and Debottam Datta spent decades around enterprise transformation before founding Trmeric. They saw the same expensive pattern across ERP, cloud, cybersecurity and AI programs: strategy begins in one room, delivery scatters across tools, and the explanation of value gets rebuilt for the next portfolio review.

Coding agents can lower the cost of creating software. They can also increase the number of initiatives an enterprise launches before improving the discipline for choosing, staffing, governing and measuring them. Faster production does not automatically create better capital allocation.

Trmeric is building an operating layer for that work. Its living transformation repository connects strategy, initiatives, resources, decisions and outcomes. Tango and purpose-built agents support scoping, prioritization, resourcing, execution signals, governance, reporting and value measurement. Foundation models provide capability; Trmeric's bet is that enterprise context decides whether it survives contact with a real organization.

Trmeric says its platform supports $1B in transformation initiatives for enterprises including Veolia, Seagate and ServiceTitan. It also reports customers have redirected tens of millions of dollars from lower-return projects and cut planning cycles from months to days. Those are company-reported results, not an audited victory lap, but they show why Hitachi Ventures sees a gap between AI spending and the ability to prove what it returned.

The $6.5M will go toward product innovation, customer expansion and scaling the business. That gives Trmeric room to turn work with more than 12 CIO design partners into a system enterprises can use repeatedly, not another transformation deck that expires before the program does.

The enterprise AI market has plenty of tools promising more output. Trmeric is working on the harder handoff: helping leaders decide which output deserved the budget, whether the work stayed connected to strategy, and where the value finally showed up. As AI makes new initiatives easier to start, that record of decisions and outcomes becomes part of the infrastructure carrying the portfolio forward.

Tensor Machines

Eight GPUs can share a model number, sit in the same rack, and settle into different answers once workload, power, memory, cooling, and hardware history start negotiating.

That variation is expensive because capacity is sold on what infrastructure can sustain, not what a spec sheet says it might reach for a moment. Tensor Machines is building a physics-informed optimization layer to make that difference measurable and useful.

The company has raised a reported $1.5M Pre-Seed round led by Omni Ventures, with Reinforced Ventures, Avesta Fund, and DraperU Ventures participating. The capital will support its open-source GPU benchmark and the models it is developing to predict how much useful work a device can sustain under changing operating conditions.

Founder and CEO Muneeb Rasool leads product and commercial strategy. Co-founder and CTO Mohamed Abdalla leads technology and engineering. Their backgrounds span Meta Reality Labs, electrical engineering, embedded firmware, industrial sensing, and machine learning.

That combination matters. Tensor Machines is not treating GPU optimization as a software setting floating above the machine. Its benchmark connects workload performance with power draw, temperature, clocks, and recovery. The physics-informed model is meant to estimate the throughput, latency, energy, and thermal outcomes available from a supported configuration or placement.

For an operator, the question is brutally practical. When another watt raises energy use without adding useful throughput, where should the workload move? When memory movement or temperature constrains execution, which operating change recovers capacity? When two devices age differently, how long can each keep doing productive work?

The investor group fits that terrain. Omni Ventures focuses on the machine-readable physical world; Reinforced Ventures on physical intelligence; Avesta Fund on constrained compute and energy; DraperU Ventures on Pre-Seed frontier technology.

The round does not answer whether Tensor Machines can predict these outcomes accurately at commercial scale. The company has not disclosed customers, revenue, fleet size, or audited performance gains, and its current models remain in private beta.

What $1.5M buys is a chance to turn a stubborn infrastructure truth into an operating product: installed compute and useful compute are different numbers. Muneeb Rasool and Mohamed Abdalla now have to make the gap visible early enough that an operator can do something about it.

Arena

A debugging agent can announce victory while the developer is still staring at broken code. Arena is building a business around the distance between that sentence and the work it claims to describe.

The San Francisco company announced a $200M Series B at a $3.1B valuation on October 8, co-led by Lightspeed Venture Partners and Khosla Ventures. Alongside the financing came a preview of its Alignment Index, measuring failures that leave evidence in an actual human-agent conversation.

CEO Anastasios Angelopoulos, CTO Wei-Lin Chiang and executive chairman Ion Stoica founded the company after its beginnings as a UC Berkeley research project. Their original platform let people compare anonymous model responses. The user supplied a preference; the leaderboard gave the rest of the industry something to argue about.

Tool-using agents have made the assignment more demanding. A useful answer can lead to a file change, a sequence of commands or a report that somebody intends to act on. Whoever delegates that work still needs to know whether the agent stayed inside the request and whether its account of the result matches the record.

Arena's new index examines unauthorized actions, false attribution and deceptive completion across 27 models and 90,000 real-world agent sessions. Those are narrow, observable signals. A high score cannot certify everything a system might do, but the definitions give an operator something concrete to inspect when an agent says it finished.

The company says it has exceeded $100M in annualized revenue and logged 7M Agent Arena sessions in less than five months. Annualized revenue is a run-rate measure, rather than a full year's collected revenue.

Salesforce Ventures, 01 Advisors, Dell Technologies Capital and Endeavor Catalyst joined the round, with support from existing investors including a16z and Felicis. Arena previously disclosed a $100M Seed and $150M Series A. The latest financing brings those announced rounds to $450M.

There is a delicate relationship inside that business. AI labs need the evaluation data, and their models appear on the public scoreboard. Salesforce Ventures points to published methods and openness to scrutiny as reasons for backing Arena. The customer buying better feedback also needs the evaluator to remain credible with everyone reading the result.

As Arena widens its measurements, a developer can bring a more pointed question to the next model decision: show me the work behind the completion message. The conversation now has a record to follow.

Hadrian

Security teams do not lose time because they lack vulnerability alerts. They lose it deciding which alert deserves a human before an attacker turns it into evidence.

Hadrian raised $40M to make that decision loop continuous. Forgepoint Capital International and SmartFin co-led the round, with HV Capital, Motive Partners, Picus Capital, and Oetker Ventures participating. Hadrian says the financing brings its total raised to $65M.

Rogier Fischer and Olivier Beg met online as 13-year-old white-hat hackers. They later founded Hadrian in Amsterdam with Maurice Clin, turning the attacker mindset they learned by hand into software that can keep probing while a security team sleeps, ships, acquires companies, and adds another forgotten subdomain.

Enterprise security has an attention problem disguised as a data problem. A scanner can produce another finding in seconds. The expensive work begins when somebody has to prove whether an exposure is reachable, exploitable, connected to something valuable, and still open after the fix.

Hadrian splits that work across two products that share context. Atlas continuously maps the external attack surface and validates which exposures can become real attack paths. Nova runs deeper, scoped agentic penetration tests on demand. Humans set the mission and make the consequential decisions; the agents handle the repetition and exploration between them.

Prash Somaiya, Hadrian's CTO, leads the technical organization carrying that model into a market where speed without boundaries is a liability. Offensive AI has to be persistent enough to think like an attacker and disciplined enough to remain useful to a defender.

Hadrian reports more than 300 enterprise customers. It also reports 10x greater visibility into critical exposures, 80% faster remediation, and 5x ROI compared with manual pentesting. Those are company-reported outcomes, but they reveal the buying argument: fewer findings competing for attention, backed by evidence a team can act on.

The $40M will fund engineering and research while expanding Hadrian across EMEA and the United States. That puts Rogier Fischer, Olivier Beg, Maurice Clin, and their team into a well-funded race with vendors pulling exposure management, vulnerability validation, attack simulation, and pentesting into the same operating layer.

The category will move toward the system that can show a security team which path is real, preserve the proof, and return after remediation to see whether the door actually closed.

Serval / Ensignia

Jake Stauch and Alex McLeod have spent Serval's short life turning IT teams into builders. Buying Ensignia brings a builder of software trust inside the same company, close to the automations those teams are being asked to approve.

Serval announced the acquisition on October 6, 2026, and appointed Ensignia founder and CEO Sam Stewart its Head of Security. Financial terms were undisclosed. Stewart previously built infrastructure security at Brex, Cruise and Culture Amp. Ensignia focused on verifying code, dependencies and build pipelines.

His new remit stretches across detection and response, corporate security, application security and infrastructure security. It also reaches the permissions, guardrails and audit trails governing how Serval's automation agents interact with customer systems. An employee's request can end in a changed account, a new entitlement or an action inside another department's tools. Someone still owns the decision that allowed it. The staff who publish those workflows remain accountable for them.

Stauch, Serval's CEO, and McLeod, its CTO, worked together at Verkada before founding Serval in 2024. They built Serval to help IT automate the work in its queue. The platform now spans IT, HR, finance, legal and security.

Catalyst builds code-based workflows from natural-language prompts. Admins review and publish them; the help desk agent uses those approved tools. Serval says actions execute deterministically through workflows admins expressly authorize. Buyers can inspect the route from request to permission to execution.

Ensignia worked further upstream, verifying the dependencies and pipelines behind production software. Its current website says it has closed its doors and points visitors to Serval. The announcement leaves the acquired technology and customer-transition plan unspecified, so the useful connection is the expertise Stewart brings to the people building that execution layer.

Serval raised a $75M Series B led by Sequoia Capital in December 2025 at a $1B valuation. That earlier capital funded an expanding platform; the price of Ensignia remains private. Serval names Fox, Live Nation, Notion, Legora, Clay and Mercor among its customers.

The next workflow an IT team publishes gives Stewart a different stretch of the trust chain to examine. He built Ensignia around how software gets shipped; at Serval, the customer also needs a defensible account of what that software was allowed to do inside its systems.

Nous Research

Open source is easy to admire before procurement asks who carries the pager.

A developer can clone an agent, swap the model, inspect the code, and decide what stays on the machine. A company has to ask who secures it, what it remembers, where the data travels, and who answers when automated work reaches production.

Nous Research has raised a $90M Series B at a reported $1.5B valuation to bring those worlds together. Robot Ventures led the round. The company named NVIDIA, Microsoft's M12, Samsung, Union Square Ventures, Y Combinator, and Menlo Ventures among the investors. TechCrunch and Axios also reported 1789 Capital.

Nous was formed in 2023 by Jeffrey Quesnelle, Karan Malhotra, Ryan Teknium, and Shivani Mitra. CEO Dillon Rolnick leads a company that earned attention by doing much of its work in public: post-training open models, publishing research, and shipping Hermes Agent under an MIT license.

Hermes has been cloned more than 24M times and drives roughly 2.5% of global AI token usage by Nous's internal estimate. Those figures are company-reported, but they explain the leverage behind the round. Nous did not have to invent enterprise interest in a conference room. It arrived with developer distribution already attached.

Hermes for Businesses will let companies customize agents, choose models by cost and performance, control more data, and build institutional memory. Its hiring says it plainly: security, infrastructure, forward-deployed engineering, support, analytics, enterprise product management, and mobile.

That list is where an open-source project starts becoming an operating company.

Enterprise buyers are purchasing the ability to change models without rebuilding a workflow, inspect what an agent is doing, control where knowledge accumulates, and avoid turning one vendor's convenience into another decade of dependency.

The $90M gives Nous room to build that layer. It also raises the standard. The company argues that AI becomes safer through auditability and competition rather than gatekeeping. Business customers will measure that claim against access controls, deployment reliability, incident response, compliance, support, and the unglamorous question of whether somebody picks up when production breaks.

Nous built its reputation by giving users more control over intelligence. The enterprise chapter will be written in how much of that control survives contact with procurement, security, and the people responsible for Monday morning.

Harness / Augment Code

Matt McClernan is handing Augment Code's software factory to a company built around getting software released. The CEO says Harness can deliver these capabilities more effectively than Augment could independently. That is a consequential choice about where the work belongs.

On October 8, Harness announced it had acquired select Augment Code assets: Cosmos, Auggie CLI, the Code Context Engine and related technology. The team behind those products will join Harness. The purchase price is undisclosed, and the announced scope is specific assets rather than the entire company.

Augment co-founders Igor Ostrovsky and Guy Gur-Ari started the business in 2022. Ostrovsky, its CTO, helped build a system that understands the relationships inside a codebase. McClernan is placing that work beside the delivery platform led by Harness co-founder and CEO Jyoti Bansal, where tests, deployments, security findings and operating costs become part of the conversation.

Cosmos coordinates agents that plan changes, write code and tests, open pull requests and respond when checks fail or reviewers ask for revisions. Engineers keep the checkpoints that require judgment, including the final merge. The useful output is a change somebody can evaluate, with the routine corrections already moving through the workflow.

The more interesting ambition concerns what those agents can learn before they start. A source repository can explain how a service is written. Delivery evidence can show which version reached production, which checks it passed and what trouble followed. Bansal plans to connect Augment's code context with Harness's Software Delivery Knowledge Graph so that experience can inform the next proposed change.

That relationship gives engineering leaders a practical reason to care about the acquisition. A vulnerability found downstream could return to the coding workflow with evidence attached. The person deciding whether to approve a correction would have a clearer connection between the original finding, the proposed fix and the checks it survived. Fewer people would have to rebuild that account by moving between tools.

Harness Cosmos is available now. The broader context and governance connections are described as integration plans, and Harness says customers can keep their preferred coding tools. McClernan's decision brings the builders closer to the release machinery; the next production finding still has to make its way back into a change an engineer is willing to approve.

BlueMatrix / Aiera

A bank can pay analysts to write research and still lose sight of where their ideas go once a client asks AI for an answer. BlueMatrix wants the firms producing that work to stay connected to its use, even when the report is no longer what the client reads.

On October 8, BlueMatrix agreed to acquire Aiera, with backing from its majority shareholder, Thoma Bravo. Financial terms are undisclosed. The companies expect a Q4 2026 close, subject to customary conditions and regulatory approvals.

Patricia Horotan and Skye Hauptman founded BlueMatrix in 1999, when getting investment research onto the web was its own undertaking. Horotan now serves as founder and CSO; David Nable became CEO in March. The acquisition announcement puts BlueMatrix's reach at more than 1,000 financial institutions and over 7,000 analysts. That is a company-reported network with a considerable interest in how research travels.

Aiera co-founders Ken Sena, CEO, and Bryan Healey, CTO, have been building the access side. Their platform carries proprietary financial content into the tools institutions already use, with permissions, attribution and a record of consumption. Fin Capital and Flybridge Capital Partners are among Aiera's existing backers, alongside a consortium of research providers.

These companies have already worked together. Their 2023 partnership brought Aiera's event coverage into BlueMatrix's authoring tools. In 2025, Aiera added transcript excerpts with playable audio. An analyst could put a management remark inside a report while keeping the reader close to what was actually said.

The acquisition takes that relationship further into AI consumption. BlueMatrix's proposed Research Provenance Standard attaches rights and attribution to structured research and calls for usage visibility back to the publisher. Aiera's API and MCP delivery gives institutional clients ways to bring content into their own workflows. The combination would put more of that journey under one company, with Nable and BlueMatrix CTO Donald Prodehl on one side and Sena and Healey on the other.

Thoma Bravo partner Adam Solomon points to demand for research institutions can license into AI systems. For a broker, the useful relationship continues after a document leaves the publishing system: whose analysis informed the answer, who was entitled to use it, and whether the firm can see enough of that activity to keep investing in the people producing it.

Automation Anywhere / Boost.ai

Nordic Capital helped Boost.ai expand beyond the Nordics. Its proposed next owner sells the software that runs enterprise work, putting the voice on a customer call closer to the people responsible for what the company does about it.

Automation Anywhere agreed on October 7 to acquire boost.ai from Nordic Capital. Financial terms are undisclosed, and closing is expected in Q4 2026, subject to customary regulatory approvals and closing conditions. This is an agreement to buy the business.

Mihir Shukla, Automation Anywhere's co-founder, CEO and chairman, wants customer, employee and business operations connected. The company acquired Aisera in late 2025 for enterprise knowledge and employee service. Boost.ai adds a customer-facing voice and conversational business built in regulated industries.

The Norwegian company began in 2016 with Lars Ropeid Selsås, Hadle Ropeid Selsås and Henry Vaage Iversen. Jerry Haywood is now CEO, Rasmus Hauch is CTO, and Henry Vaage Iversen remains CCO and co-founder. Nordic Capital invested in 2021, supporting international expansion before this proposed handoff to an automation operator.

Boost.ai already supports authentication, data retrieval and approved transactional actions through integrations. Its voice software can use speech-to-speech for lower-risk exchanges and a structured pipeline for regulated transactions. That choice makes the authority behind a fluent answer part of the product itself.

A service team buying that voice has company. Process owners have to decide which actions it may perform, security teams need to understand the data it touches, and employees need a route into the conversation when judgment is required. Adding more connected systems makes those relationships more consequential.

Ankur Kothari, Automation Anywhere's co-founder and COO, describes the intended flow from an expressed need through company rules, people and systems to execution. Haywood wants the conversation connected directly to the work. Their shared ambition gives a service buyer a larger implementation conversation than choosing an agreeable voice.

Boost.ai reports 600+ live AI agents and 150M+ automated conversations on its company page. That page gives no annual period for the conversation count. Scale supplies experience; customers still need evidence for the particular actions they intend to authorize.

The next procurement discussion can follow a caller's request all the way into the systems allowed to change the answer. Somebody in that room owns each permission.

Vates

For 10 profitable, self-funded years, Vates financed its speed with revenue. On October 6, the Grenoble company made a different choice: take €30M, to move faster while its market is being rearranged around it.

IRIS invested through its Growth fund and Bpifrance through Large Venture. This is Vates' first institutional round. Co-founders Olivier Lambert, Nithida Vialle, and Julien Fontanet remain majority shareholders, which matters because control is part of what the company is selling.

Vates builds an open-source virtualization stack centered on XCP-ng and Xen Orchestra. One runs the virtual infrastructure. The other manages, orchestrates, and backs it up. Customers can use the same software whether they buy support or build from source.

That architecture became more commercially interesting after Broadcom acquired VMware. Licensing changes, packaging decisions, and cost increases pushed infrastructure teams to revisit choices that once looked permanent. A migration is still painful, but staying put acquired a price tag and a governance question.

Vates reports more than 1,000 customers across about 100 countries, over 65% growth in 2025, 90% of revenue outside France, and North America approaching half of its business. Those company-reported figures describe a business raising to compress time, not to discover whether anybody wants the product.

The capital will fund performance, storage, security, automation, and large-scale infrastructure for newer workloads such as AI. Vates will also expand commercial teams and its partner network.

The partner piece may be the least glamorous and most important. Someone has to map workloads, preserve backups, test dependencies, train operators, and answer the phone when a migration meets production reality.

Vates must add that machinery without turning open source into a decorative label. The founders say XCP-ng and Xen Orchestra will keep their licenses, support will remain the business model, and founder control will hold.

The capital creates a faster route into accounts reconsidering VMware, and makes Vates' promise harder to hide behind. Buyers will judge what remains open, how migrations behave, and whether support can carry critical infrastructure after procurement ends.

Vates now has the money, timing, and control structure to offer Europe an infrastructure answer from Grenoble. The next part will be written inside data centers where reversibility has to work before anyone calls it sovereignty.

Ultra

Ultra gave OP1 wheels that lock. The robot can move between work areas, then stay put and pack orders from a wall outlet. A warehouse already has enough moving parts without making its packing station another one.

That choice sits inside a $50M Series A led by Framework Ventures, announced October 9. Ultra's $62M funding total includes an earlier $12M seed led by Y Combinator and NextView, according to reporting. Y Combinator also participated in the Series A.

Brooklyn-based Ultra was established in 2024. CEO Jon Miller Schwartz, CTO Oliver Ortlieb, and COO Max Friefeld previously built Layer By Layer and Voodoo Manufacturing. Co-founder and Chief Scientist Chetan Parthiban brings applied AI experience from Arena AI, a separate company from the evaluation platform featured earlier in this issue.

Their new customer is buying useful work inside a warehouse that is already running. OP1 handles repetitive packing, sorting and kitting. Ultra's site shows installations in New York, Atlanta, New Jersey and Austin, with different packaging workflows across those locations.

The business model makes that operating relationship hard to escape. Customers pay an integration fee and recurring charges for hardware and software support, according to reporting. Each invoice arrives while the machine is still expected to earn its place on the floor.

For the buyer, a monthly service arrangement can make automation easier to approach. For Ultra, it keeps installation, repairs and the customer relationship close to the product. A machine delivered is the beginning of a service obligation. The rental decision gives the customer a continuing reason to judge the service.

Physical Intelligence supplies robotic AI models used in Ultra's deployments. Ultra concentrates on the hardware and field operation. Those roles connect model development to actual customer work, where a packing mistake still has to be corrected by somebody.

Ultra says its robots have packed more than 500,000 orders. That cumulative, company-reported count puts real activity behind the story; public revenue, fleet size and support margins remain undisclosed.

Framework's capital backs expansion of the robotic fleet. The interesting commercial question travels with every installation: how much continuing work does it take to make a robot a dependable part of somebody else's business?

Jon Miller Schwartz and his co-founders are carrying that question from one loading dock to the next, with customer orders continuing to arrive while the fleet grows.