16 Companies Making Agents More Useful Are Raising Funding

@Guo Yunxiao

In less than a month, the debate over whether OpenClaw itself was merely a wrapper came to an end.

Fortunately, many also realized that OpenClaw offered far more than wrapping. Or rather, OpenClaw was a wrapper — a framework that made models truly usable: gateways, heartbeat mechanisms, Skills libraries, memory files. The lobster wasn't the point; the thinking behind this framework was. So Silicon Valley distilled a new term: Harness — optimizing everything around the model so that agents can actually get work done.

For a long time, entrepreneurs leveraged models while simultaneously dodging and resisting them, comforting themselves with "wrappers have wrapper value." This was the first time the definition of "wrapper" finally showed the possibility of transcending model value itself.

Over the past month, elsewhere spoke with numerous entrepreneurs and investors. One overarching impression: OpenClaw is too narrow, while harness is too broad — when we talk about these two terms, we don't actually know what we're really discussing. Consensus fragments across sub-directions.

If we must find a greatest common denominator, it's probably: built for agent — designing everything natively for agents.

Following this thread, we gathered, organized, and experienced some recently launched products or companies currently fundraising, mapping out directions and companies we find interesting. Broadly: agent computers, A2A networks, agent identity and payments, multi-agent collaboration, and solving agents' context acquisition and long-term memory.

This batch of startups has already reached valuations around $100 million, with nearly all actively raising new rounds.

Beyond these, there are directions we find curious but haven't expanded on here. Skills marketplace, for instance, where major tech companies seem more enthusiastic. Agent security is an important proposition, but we haven't yet learned of new developments from Chinese teams. Qi Junyuan's Cijian Wuxian secured $10 million from IDG Capital and StepFun to build a GUI agent phone, which implies a larger discussion about how GUI interfaces might evolve.

After reading this first-hand inventory, you should quickly grasp what the lobster world has been watching and anxious about lately. Another interesting note: not a single one of these companies has a Chinese name — they're all English.

Agent Computer

Hardware Container

Agents need a vessel; this is already consensus forged through lobster practice. The cloud can only be a crippled version, since there's neither context for it to learn from nor "things" for it to operate; on your primary machine, data is rich and tasks can extend further, but data security becomes a problem, and "it fights you for the computer" while you work — shut your laptop and it goes offline too. A spare machine may be the best option, which has directly driven Mac mini prices to double market rate. The agent computer emerged. Literally: a computer specifically for agents.

Pamir AI, Violoop, and Tiiny AI

These are among the faster-moving fundraisers, with valuations all approaching the $200 million range. One investor who has looked at these companies said: "(Recently it's been) doubling every week." For reference, a February 2026 report noted Pamir AI's valuation at approximately $25 million; Violoop completed two rounds within one month; Tiiny's fundraising is also advancing rapidly.

This feels like a continuation of the last wave of consumer electronics AI hardware investment enthusiasm.

Hardware prices generally fall in the $200-300 range, but each has distinct characteristics. Pamir has Claude Code built in, with hardware providing a 7×24 runtime environment and USB peripheral connectivity; Violoop connects via HDMI to computers, with on-chip screen viewing and simulated keyboard-mouse operations, while inference goes to the cloud; Tiiny uses the heaviest specs — 80GB memory plus 190 TOPS compute, with models and data kept entirely local.

This is indeed an ideal battlefield for demonstrating "hardware-defined capability." PC manufacturers, model vendors, traditional peripheral makers, AI hardware companies, even ordinary 3D printing enthusiasts — all can participate.

Some believe the direction should be lighter and more portable, shrinking to card-sized or smaller; others aestheticize it as part of desktop decor, even IP-ifying it with screens or holographic projection. Some think input should simply plug into IM, keeping hardware light; others equip it with voice input buttons, even a full keyboard.

Faced with this flourishing diversity and the lofty valuations of leading players, investors are also looking upstream at agent chips — though this isn't entirely a startup domain.

Recent changes have also affected another batch of consumer electronics AI hardware that originally tried to build agents as standalone offerings. Whatever their hardware story was, they're now pivoting toward personal context entry points, becoming one link in the broader chain. According to our understanding, multiple projects currently fundraising have adjusted their valuations.

Agent to Agent Network

Among OpenClaw's earliest use cases, Moltbook was the most eye-catching.

This was a social network where only agents could participate and interact, product-wise resembling a BBS. Despite controversies ranging from hype exceeding actual content to humans posing as agents to post, it was acquired by Meta in extremely short order. This phenomenon-level product aptly demonstrated OpenClaw ecosystem potential: through agents publishing and browsing, human-to-human matching was accomplished — a prototype of stranger social networking.

Second Me

Most companies introduced in this article were founded after agent became consensus direction, making them a different generation from Second Me. But the fastest mover domestically was still Tao Fangbo.

Second Me had long sought to create each person's "digital twin" as their interface to the digital world. Because these "digital twins" existed, Second Me was able to build out its social network in extremely short time. After going viral, Second Me rapidly organized several A2A hackathons, trying to discover what might grow on this network.

The company was founded before ChatGPT's release. After several pivots, it decided to participate in building the agent ecosystem. Before these, Sequoia Capital China, Linear Capital and others were its investors. They have also been in contact with some investment institutions recently.

Elys

This was another product that blew up during Spring Festival, from Natural Selection, which had previously launched EVE. On the user side, they skipped past the agent deployment part, doing only agent social networking. In early 2026, Natural Selection announced $30 million in new funding from Alibaba, Ant Group and other institutions — at which point the Elys project had not yet launched. According to our understanding, Elys is just one of the team's experiments.

A2A social networks have consistently faced much controversy, with the dominant critique being "performance for humans." In fact, this is a problem of the social network form, not of A2A networks themselves.

There are also products attempting to connect agents from other angles.

EigenFlux

This is a broadcast network enabling large-scale communication among global agents, now open-sourced. Once connected, your agent can broadcast needs or capabilities to the entire network, and also subscribe to and receive broadcasts of interest using natural language. Their logic: broadcasting is a more suitable information acquisition method for agents — one-to-many, one-shot, token-saving. In their most recent version update, we received an official DM "looking forward to further communication with investor agents."

Network effects may be the story that excites investors most. A2A theoretically has network effects too. Every additional node connected increases the value of all nodes — which is why A2A networks, even when they currently look "useless," still warrant attention.

Identity / Payment

Identity and Payment

The further corollary of A2A networks is value flow within the network.

The most typical scenario is agent-to-agent hiring. The assumption: when your agent has learned and iterated on your skills and cognition, it can complete tasks other agents cannot, at which point your agent may be discovered and hired within the network. This is not a new concept — many entrepreneurs have envisioned building such markets.

TTC is an interesting company. TTC stands for True Talents Connect. Two months ago, they raised a Series A led by Houxue Capital, with current valuation around $100 million. This is an AI headhunting company with a CTO, doing both traditional human recruiting and attempting agent-to-agent hiring matching.

The less mature piece is payment. To have agents do more for you, especially reaching into corners of daily life, payment is a problem you'll eventually face. Existing payment systems are designed for humans; all risk control measures are difficult for agents to cross.

Payments have always been good business; agent payments are too, and it's an unformed large pie. Startups and global giants are competing on the same field again.

FluxA

FluxA was founded by a team with former Ant Group background, designing an entire payment system native for agents: Agent Wallet, AgentCard (virtual card), and their self-developed AEP2 embedded payment protocol, with stablecoins and x402 protocol at the底层. They broke out through a "lobster red envelope grabbing" campaign, having lobsters send and grab red envelopes from each other, achieving significant agent user growth. Currently, their valuation is in the tens of millions of dollars.

Clink

This company takes the inclusive finance route: its Agentic Payment Skill lets agents complete payments through credit cards, Apple Pay, and local wallets, and has obtained PCI Level 1 certification. Their logic — the dividends of the AI era shouldn't belong only to those with crypto wallets; freelance workers in small towns should also be able to let agents spend money through the most familiar methods. Several on-demand AI products have already integrated Clink. According to our understanding, Clink has completed a multi-million dollar seed round led by BV Baidu Venture Capital and others.

Linked

This project emerged from an A2A hackathon, supports Second Me login, and its creator Sol is only 13 years old. This product takes a broader approach, positioning itself as a Human × Agent identity and reputation network — integrating identity, wallet, marketplace, multi-agent collaboration, and six-dimensional reputation scoring on one platform. The logic: for agents to transact with each other, they first need to know if the other party is reliable.

Regarding payments, there is a divergence between crypto-finance and traditional finance routes. The crypto route is natively aligned with agents, but for various reasons has never achieved sufficient global penetration; traditional finance has more robust infrastructure, but turning all human-facing systems toward agents is a massive engineering undertaking. Beyond these, some believe the timing for payments isn't ripe yet, since agent networks and high-value tasks are not yet ready.

This direction is still very early, but the logic is clear: every shift in economic actors on the internet has been accompanied by payment paradigm updates — from Visa's card networks to PayPal's digital wallets to Stripe's API payments — each not merely a technical upgrade but a change in commercial participant identity. This time, the new participant is the agent.

Multi Agent Workspace

Multi-Agent Collaboration

In productivity scenarios, multi-agent interaction is almost another matter entirely.

From the human perspective, one person directing multiple agents to work is necessary — context and attention are both bottlenecks for single agents; in organizations where everyone has their own agent, how to organize these agents' different permissions, contexts, and collaborative relationships is also a problem. The two converge on the same need: a multi-agent collaboration method.

This is a proposition naturally belonging to enterprises.

Collaborative office software is long-established business, with giants everywhere. But adding AI to old systems is merely improvement; agents need spaces designed natively for them. This is certainly also a harness. Building peripheral systems for single agents is hard to establish lasting moats, but the coordination layer among multiple agents involves organizational relationships, permissions, trust — things models cannot consume.

Floatboat

Founder Tan Shaoqing's logic: agents must first stay where you work to have things to record, so files, web pages, and agents are designed as three parallel windows, with production scenario context all inside. At the collaboration level, the team open-sourced two protocols: one is Selfware, an inter-agent file transfer protocol, attempting to let agents under different frameworks pass files and receive tasks from each other; the other is IACT, turning agent-presented options from typing into clickable buttons, reducing interaction friction. They recently received a seed round from Sequoia Capital China and WeLight Ventures.

Moxt comes from the original Motiff team, from two colleagues initially exploring "collaborative lobsters" to the entire team abandoning their old product to join Moxt — five days. The team believes AI agent's native workspace should be: all content converted to .md, .csv, .html, with directory structure as file system — these knowledge organization methods most familiar to AI during training. Based on this, they made a product even they cannot rigorously define, Moxt, with the apt answer possibly being "Obsidian + Claude Code + cloud collaboration."

Clawith comes from BISHENG, not strictly a new startup either. This product takes the open-source route, positioning itself directly as "OpenClaw for Teams." Beyond giving each agent persistent identity, memory, and independent workspace, they upgraded the lobster heartbeat mechanism: from an "alarm clock" waking every 30 minutes to a system called Aware — continuous perception where agents set their own triggers, adjust themselves, cancel themselves, sensing continuously and acting on demand. When agents enter organizations, they need to know who is boss, whose instructions carry more weight, what colleagues' agents are each busy with — this is their other innovation, the Relationship graph.

Multica This is a product we discovered just hours ago, from Zhang Jiayuan of Devv.ai. Their approach turns task management into a shared interface for humans and agents — form resembling Linear, but agents are first-class citizens. The product's key may be using kanban to assign tasks to agents, manage progress, then transform into team-shared reusable skills. The product is already open-sourced. ToB software is a赛道 investors are very cautious about, for obvious reasons. The wreckage of the last SaaS wave, AI's massive impact on enterprise software in the US market, market share and user habits already captured by large company products, and the most oft-cited Chinese enterprise service market soil — all are reasons for disbelief. But those who see opportunity think, "those who redesign interaction for touchscreens will defeat those who shrink websites to fit phones." Worth noting: besides ByteDance-background Zhang Jiayuan, other key figures in these products skew older: Tan Shaoqing is post-80s, on his third entrepreneurship; Moxt described themselves in retrospect as "a group of aging 'mid-olds' and 'old-olds'"; Clawith news largely comes from BISHENG co-founder Qin Rui, who is 34 this year. This may be cause and effect with this赛道's particular nature.

Context / Memory

Context and Memory For agents to do good work, they must both see the present and remember the past. Context solves "what should be seen right now" — grabbing information needed for current tasks from your files, browser, chat records, and stuffing it into the agent's limited context window. Memory solves "what was learned last time" — remembering your preferences, work habits, and accumulated experience across sessions and tasks. Neither is new, but both are core parts of harness.

Context's problem is that it's hard to exist as a standalone product. Context is ephemeral, gone after use, unlike memory that accumulates thicker over time. And models' own context windows are rapidly lengthening, with agent products increasingly building context management as native capability — compression, filtering, tiered loading, these tasks increasingly don't require external services.

The most famous overseas company in this direction was formerly Rewind, which offered full life recording on Mac + iPhone, skyrocketing in valuation after emerging, then pivoting to hardware product Limitless, and ultimately selling to Meta early this year.

This is almost the microcosm of Context as a whole: standalone context capture tools can accumulate data, but struggle to turn data into value users will pay for — pure recording has too low usage frequency, always behind the scenes, users don't notice it.

AirJelly This is a newly beta-launched product, whose predecessor was the open-source project MineContext by post-00s founder Bai Te during his ByteDance tenure; upon deciding to start a company, Wuyuan immediately invested. From MineContext experience he reached a key judgment: full-screen recording has limited value, intent recognition matters more. Consequently, AirJelly also does Context's next step: Proactive — thinking of and directly helping you complete the next task.

Beyond this, another route is open-source local. Because of context data sensitivity, some domestic teams believe relevant functional modules should remain entirely local.

Memory is not a concept the lobsters popularized; its heat came earlier. After the 2024 RAG wave, "making AI remember things" already became an independent赛道. Later OpenClaw used the most朴素 method — a few markdown files — to prove memory's experiential value.

Unlike Context, Memory has potential to form standalone products, because memory is沉淀able asset — the longer used, the deeper the accumulation, the higher the migration cost, naturally carrying compounding effects and user stickiness. And memory is extremely scene-dependent; general model vendors can hardly do well in every vertical scene, giving third-party memory services room to exist. For this reason, startups in this direction are numerous domestically and overseas.

MemU Founded by former ByteDance Seed team member Chen Hong, core concept is "Memory as File System" — turning memory into visible, controllable, organizable file structure rather than vector black box. He chose to切入 from emotional companionship, as this is where user memory dependency is strongest: you must remember every bit of the user, otherwise it's not "companionship." MemU also introduces Theory of Mind, expanding understanding atop memory.

MemOS (Memory Tensor) Directly targets enterprise scenarios, doing a "governable memory operating system" — not just letting AI remember things, but making memory auditable, rollbackable, migratable, cross-platform shareable. They even plan to launch a "memory marketplace," letting developers package enterprise knowledge into downloadable "memory units" for上架. Last year, they completed an angel round of nearly RMB 100 million, with investors including Futeng Capital, CICC Capital and others.

At this layer, related startups all face a common problem: how to create user perception.

Context and Memory are "invisible infrastructure," with effects only gradually showing after extended use. This is likely also the challenge more other agent infra faces — this thing is important, but is it a module in someone else's product, or a "xx-direction enhanced" agent that can complete the next step? Both stories seem somewhat awkward, unless there exists an open-source agent framework that can accommodate these functional components.

As it happens, OpenClaw streaked across like a meteor.

Cover image: Pieter Bruegel the Elder, The Tower of Babel, 1563, Kunsthistorisches Museum