In one sentence: Work Agent—a "proxy-play" AI with persistent memory, resident in chat channels, wielding system-level permissions, and growing through a skill ecosystem—converged this summer of 2026 via three independent tracks: grassroots open source (OpenClaw), startup (Hermes), and big-tech incubation (WorkBuddy), achieving product-market fit.
To judge whether an AI category is valid, don't look at the Feature List—look at whether User Stories are a top-level entry in the official docs. When "what others have done with it" persuades potential users more than "what it can do," the category has crossed from Demo into daily infrastructure.
Look at OpenClaw, Hermes Agent, and WorkBuddy together, and you'll find they share the same skeleton—and all clearly differ from earlier chatbots or pure coding agents:
OpenClaw's growth curve is itself the best footnote to this category's explosive speed: after two renames (Clawdbot → Moltbot → OpenClaw), the project was officially named and released in late January this year, and within a week its star velocity ranked among the fastest in GitHub history, reaching 346k+ stars by early August 2026; update (2026-09-13): live repository data has risen to 389,554 stars and 81,888 forks.
But explosive growth also brought security debt along for the ride:
How fast a category runs, that's how much security debt it accrues—these two things have always been two sides of the same coin.
Update (2026-09-13): the project didn't stand still paying off debt—OpenClaw 2.0 (v2026.8.1) released on August 30 is the largest update in the project's history, merging 16k+ PRs with 933 contributors, adding explicit session permissions, pre-install plugin trust review, credential desensitization prompts, and other mechanisms—an official concentrated repayment of the security debt above, though far from "paid off."
State lives on your machine,
not a vendor cloud. — OPENCLAW · The self-hosting selling point repeatedly emphasized on the official site
The official Hermes User Stories page doesn't discuss model parameters—it's all about "what people are actually using it for," roughly divided into five categories:
| Usage Category | Real Scenarios |
|---|---|
| Personal Assistant & Task Management | Aggregate inbox and push to Slack; sync tasks across Obsidian and Apple Calendar; schedule in natural language ("every weekday at 9am"); proactively push reminders via Telegram / Discord / Signal; leverage persistent memory to recover context across sessions. |
| Developer Workflow | One user said that after ten days the Agent "knows my codebase better than I do"—it has internalized file priorities, coding preferences, deployment pitfalls, and common API call sequences. Advanced usage involves multi-agent orchestration: one plans, one writes code, one runs QA, with the collaboration process monitored and debugged in real time. |
| Enterprise Business Operations | Integrate with ticketing systems like Plane.so for automatic triage and dispatch; build a "Chief of Staff" master Agent coordinating multiple project sub-agents with independent memory layers; use for sales lead generation, CRM integration, batch-generating Slides / Sheets. |
| Research & Content | Automated research stack periodically scrapes AI developments, aggregates HackerNews, distributes briefs to Slack, Discord, Notion, Obsidian; content creators use it for weekly topic research, voice-first podcast generation, social media curation matching personal writing style. |
| Cost Optimization | One user compressed 5-day token spend from ~$130 to $10: switching providers, tiered model routing (cheap models for mechanical tasks, expensive models only for complex problems), even running a local deployment on a $10/mo Hetzner VPS. |
The last item best illustrates that "working" is a genuinely occurring economic behavior—when users start optimizing the cost of "hiring an AI to work" the way they optimize cloud bills, this category has crossed the novelty phase and entered real sustained usage.
If Hermes's User Stories are officially curated cases, the OpenClaw community's own 50+ real use cases are wilder: someone photographed and cataloged every item in every room of their home, letting the Agent remember "where the tracker board is"; someone used photos to generate a homeschool curriculum plan, then linked the item inventory to arrange hands-on activities; someone built an overnight auto-generated MVP—creating a video app for kids with no ads, no algorithmic recommendations, and no autoplay in four days, with the Agent completing the entire process while the user slept.
The cases that really explain this usage deeply come from two "How I AI" podcast episodes. Jesse Genet uses 5 OpenClaw Agents to manage home, finances, and code, turning Agents from work tools into life infrastructure; ChatPRD founder Claire Vo's transformation is more dramatic—on her first try she wiped her family calendar clean, becoming a thorough skeptic, but later ran 9 specialized Agents on multiple Mac Minis and old laptops, each managing family scheduling, sales leads, kids' homework, podcast prep, and course operations. Her takeaway: "multiple specialized Agents beat one generalist Agent"—which is really the same thing as Ramp internally crystallizing AI usage into 350+ Skills, just replayed at the individual scale.
Ecosystem data also confirms this curve: public skill registrations on the skill marketplace ClawHub reached 13,729 by February 2026, with the community curating 5,494 into a high-quality collection, and NVIDIA stepping in to do skill security scanning (Skill Cards); update (2026-09-13): per Hermes Skills Hub's August 26 stats, ClawHub registered skills have surged to 69k+, the largest of the 11 tracked skill registries. Even more telling is TrustMRR's data: ~128–129 OpenClaw-related projects generated $281k in 30-day revenue, with the most profitable project being "helping you use OpenClaw more cheaply." Update (2026-09-13): the same TrustMRR live stats show participating projects have grown to 187, but 30-day revenue has fallen back to ~$99k—more participants, but the pie didn't grow with them.
In platform economics, the first to profit are always the tools that reduce costs for the infrastructure itself—this pattern holds once again in the Work Agent category, except that as more entrants arrive, this "cost-reduction business" itself is also being diluted.
Tencent's official positioning for WorkBuddy stops just short of saying "this is our OpenClaw"—TechNode's headline directly reads "Tencent launches OpenClaw-like workplace AI agent WorkBuddy". But it solves OpenClaw's natural shortcomings in enterprise scenarios: an open-source Agent requiring system-level permissions, installed on employee machines, with state remaining locally—this combination presents high learning and trust costs for ordinary workers, and is a nightmare for enterprise IT / compliance departments.
WorkBuddy's product logic is multi-agent orchestration + local execution: a single natural-language instruction gets decomposed into parallel subtasks—one Agent fetches public web data, one summarizes an uploaded report, one assembles a PPT, "three things happening simultaneously". Tencent Cloud's own description is "100+ built-in Experts", covering industry scenarios from market research to financial analysis, with the tagline "Not just answers. Finished work." Technically it supports MCP-style tool integration, and the model layer isn't locked to a single provider—it can route among Hunyuan, DeepSeek, GLM, Kimi, and MiniMax.
| Version | Pricing | Notes |
|---|---|---|
| Overseas | Free / Pro ~$9.95/mo / Team ~$40/seat/mo | Three subscription tiers |
| Domestic Personal Pro | ¥58/mo | Includes 2,000 credits |
| Domestic SaaS Enterprise | ¥198/seat/mo | Tencent Cloud |
| Domestic Dedicated Cloud | ¥316/user/mo | Private deployment |
What's truly interesting about this product line isn't feature parity, but that "local execution, multi-model swappable, enterprise-grade"—these three words appear simultaneously in the Work Agent category for the first time—almost as if written specifically for enterprise customers who can't directly adopt US open-source solutions due to compliance and data sovereignty concerns. It's safe to predict this won't be the only "OpenClaw-like but enterprise-safe" localized variant; in the next year, such "alternatives" will likely appear in every major market.
Validation (2026-09-13): this prediction came true much faster than expected when it was written, and all within the same month. ByteDance launched "Doubao Work" on August 25, merging TRAE and Coze into the Doubao brand with deep Feishu integration; Alibaba merged QoderWork, MuleRun, and Wukong into "Qwen Office" for public beta on August 3, and by September 4 monthly users had surpassed 30M with enterprise users exceeding half, then on September 7 launched the industry's first "Multi-person Workspace" (generate a dedicated workspace supporting 100-person real-time collaboration with a single sentence). Tencent itself didn't stop either: on September 2 WorkBuddy Open Platform (open.WorkBuddy.cn) officially launched, introducing 100+ ecosystem partners in the first batch, simultaneously releasing 9 co-branded smart hardware devices (Plaud, Rokid, Insta360, iFLYTEK, etc.) and an industry-facing "Buddy App" entry point, with the first batch connecting 30+ enterprises including Tongdaxin, GF Securities, and Tencent Health. Tencent Cloud VP and WorkBuddy lead Liu Yi stated the goal is to make WorkBuddy "the operating system of the Agent era." All three, almost within the same month, turned "Tencent alone builds an alternative" into "three giants in a dogfight."
The details of Nous Research's funding round are worth unpacking: TechCrunch reports the company is closing a new round at a $1.5B valuation, raising at least $75M, led by Robot Ventures with heavyweight follow-on from USV (Union Square Ventures) and others—another step up after previously raising $70M from Paradigm, North Island Ventures, OSS Capital, Balaji Srinivasan, and others. What supports this valuation isn't narrative, it's numbers: Hermes Agent has accumulated ~214k stars and ~40k forks on GitHub, with a commercialization path of open-source self-hosting + cloud subscription ($20–200/mo) walking on two legs. Update (2026-09-13): GitHub live data shows stars have grown to ~245k and forks to ~51k.
Put three things together and the logic of this bet becomes clearer: OpenClaw proved that a grassroots open-source ecosystem can grow tens of thousands of skills and hundreds of thousands of developers in months; Hermes proved this model can be built into a business with clear paid tiers, and earn real-money valuation backing from top-tier institutions; WorkBuddy proved this logic can be replicated by big tech into an enterprise-grade, locally compliant version.
Three different starting points (grassroots open source, startup, big-tech internal incubation) converged on the same product form—that itself is the hardest evidence that the category is valid. Not a story told by one company, but three independent tracks converging simultaneously.
| Time Horizon | Key Judgment |
|---|---|
| Short-term (Now – 6 months) | Skill/plugin ecosystem richness and security review mechanisms will determine user retention faster than underlying model capability. ClawHub's 13,729 skills and NVIDIA's Skill Cards scanning coexist with Cisco's audited 26% vulnerability rate—showing that "ecosystem prosperity" and "security debt" grow in lockstep. When procuring or building a Work Agent, treat skill review as a hard gate, not an afterthought. |
| Medium-term (6 – 18 months) | The market will clearly split into "out-of-the-box cloud Work Agents" (Hermes cloud subscription, WorkBuddy SaaS) and "self-hosted open-source solutions" (OpenClaw proper and its various localized variants). The core variable in enterprise procurement decisions will be data sovereignty and compliance costs, not feature differences—WorkBuddy's "OpenClaw-like but locally executed" positioning is the earliest product of this split, and corresponding versions will likely appear in other markets. |
| Long-term (18+ months) | When a product category's official docs treat "what others have done with it" (User Stories) as more important and persuasive to potential users than "what it can do" (Feature List), the category has crossed from Demo stage into the threshold of daily infrastructure. Hermes making User Stories a top-level doc entry, Claire Vo going from skeptic to power user running 9 Agents, the geek experiment of operating a company at $400/mo—these scattered cases, aggregated, are all answering the same question. |
When will AI stop being a "chat window" and become "the colleague who never sleeps"? The answer is no longer "whether," but "which role comes first." Three independent tracks converging on the same product form in the same summer; users starting to optimize the cost of "hiring an AI" like they optimize cloud bills; the first profitable layer in the ecosystem being cost-reduction tools—these three signals stacked together declare the Year One of Work Agents.
This article was written in late July 2026 and published on August 1. A month and a half later, the core judgments have largely held up—but the convergence speed on the big-tech track has been far more intense than anticipated when writing this piece.
The grassroots open-source and startup curves are both still rising. As of September 13, 2026, OpenClaw's repository has reached 389,554 stars (81,888 forks), and Hermes Agent's repository has grown to ~244,976 stars (~50,868 forks). OpenClaw hasn't stopped iterating either: on August 30 it released the largest-ever version 2.0 (v2026.8.1), merging 16k+ PRs with 933 contributors, adding explicit session permissions, pre-install plugin trust review, credential desensitization prompts, and other security mechanisms—an official concentrated repayment of the security debt mentioned in this article.
The big-tech incubation track has gone from "Tencent alone builds an alternative" straight to "three giants in a dogfight." ByteDance launched "Doubao Work" on August 25, merging TRAE and Coze into the Doubao brand with deep Feishu integration; Alibaba had already merged QoderWork, MuleRun, and Wukong into "Qwen Office" for public beta on August 3, and by September 4 monthly users had surpassed 30M with enterprise users exceeding half, then on September 7 launched the industry's first "Multi-person Workspace". Tencent itself on September 2 launched the WorkBuddy Open Platform (open.WorkBuddy.cn), introducing 100+ ecosystem partners, 9 co-branded smart hardware devices, and an industry-facing "Buddy App" entry point to fight back. All three, almost within the same month, turned "alternative" from a single-point validation into a head-on clash.
But the other side of ecosystem commercialization should also be seen. The TrustMRR data cited in this article (128 projects, $281k in 30-day revenue) was actually a snapshot from February–March 2026; as of September 13, 2026, the same live stats show participating projects have grown to 187, but 30-day revenue has fallen back to ~$99k—more participants, but the pie didn't grow with them. ClawHub's skill registrations likewise grew from 13,729 to ~69k (2026-08-26 count). Both curves together tell the same story: the "ecosystem prosperity" of the Work Agent category is real, but the early path of profiting from information asymmetry and first-mover advantage is being rapidly diluted by more entrants—this and the big-tech track going from "one" to "three-way melee" are really the same "category validated, then immediately enters fierce competition" story replayed at different scales.