The most important thing to understand about the four AI agent projects generating the most press coverage in mid-2026 is that they are not competing with each other. OpenClaw accumulated more than 370,000 GitHub stars by late May 2026. Hermes Agent, from research lab Nous Research, overtook OpenClaw on May 10 to claim the top spot on OpenRouter's global daily inference rankings, processing 224 billion tokens in a single day against OpenClaw's 186 billion. Genspark, the Palo Alto super-agent platform, surpassed $200 million in annualized revenue when it extended its Series B to $385 million in April 2026. And Manus, the Wuhan-founded autonomous agent, was acquired by Meta in late December 2025 in a reported $2 billion-plus deal — which China's National Development and Reform Commission ordered unwound in late April 2026.

Call any of these the winner of the agent layer and you have misread every one of them. The dominant commercial players in agents are coding-tool incumbents whose footprints dwarf this group. What the four projects above represent is four different answers to a single question: what does an agent business actually look like? The answers are barely compatible with each other — and that incompatibility is the real story of where the agent market stands in mid-2026.

Coding Incumbents Already Lead the Agent Market by Every Commercial Measure

Any serious view of the agent layer has to start with the coding-agent platforms, not the four projects generating the most social-media conversation. OpenAI reported in its March 2026 capital-raise announcement that Codex serves more than 2 million weekly users, up five times in three months, with month-over-month usage growth above 70%. Anthropic's Claude Code reached roughly $1 billion in annualized revenue within six months of launch — and in a February 2026 survey of 15,000 developers by the Pragmatic Engineer, 46% named it their most-loved coding tool, more than double the share for any competitor. Cursor, the AI code editor from Anysphere, raised $2.3 billion in November 2025 at a $29.3 billion valuation, and was reported in April 2026 to be in advanced talks for a new round at a valuation above $50 billion. GitHub Copilot, the longest-tenured of the category, sits at roughly 4.7 million paying users. Cognition's Devin anchored the autonomous-software-engineer category and extended into enterprise distribution through a Cognizant partnership in January 2026.

Compared with those numbers, the four projects that dominate agent coverage in tech media are small. What they offer instead is clarity about which business model each is testing.

Four Paradigms, Four Business Theories That Cannot All Be Right

OpenClaw is the open-source-as-infrastructure paradigm. Created by Peter Steinberger — the Austrian engineer behind PSPDFKit, which raised $116 million from Insight Partners in 2021 — OpenClaw launched in November 2025 as Clawdbot. Anthropic asked Steinberger to change the name to avoid confusion with Claude, triggering two rapid renames before the project relaunched as OpenClaw on January 30 and crossed 100,000 GitHub stars within 48 hours. It is MIT-licensed, runs locally on a user's hardware, and connects to more than 50 messaging platforms. Its sharpest public data point is the cost of running it at scale: Steinberger ran 100 agents continuously for 30 days and generated a $1.3 million OpenAI token bill — a bill OpenAI itself covered, and a number no vendor's marketing page would ever voluntarily publish. OpenClaw has no consumer revenue model; its value sits in open-source distribution and ecosystem adoption. Steinberger joined OpenAI in February 2026 to lead personal agents research, while OpenClaw transferred to an independent foundation with OpenAI as a financial and technical sponsor.

Hermes Agent is the research-lab-as-distributor paradigm. Released February 25, 2026 by Nous Research — the lab behind the Hermes, Nomos, and Psyche model families — it crossed 140,000 GitHub stars in under three months and became the most-used agent on OpenRouter by daily token volume. NVIDIA has positioned the agent as a showcase workload for its RTX and DGX Spark hardware. Hermes's most-discussed architectural feature is a skill-and-memory loop: after every task involving five or more tool calls, the agent runs a reflection step and generates a reusable skill file so it does not repeat the same discovery work in the future. An autonomous background process called the Curator grades and rewrites underperforming entries on a weekly schedule. There is no Hermes Agent subscription — the project exists to drive usage of Nous's models and inference infrastructure.

Genspark is the subscription-first agent-workspace paradigm. Founded in 2023 by former Baidu vice president and Microsoft Bing veteran Eric Jing alongside co-founder and CTO Kay Zhu, and based in Palo Alto, Genspark pivoted in April 2025 from personalized search to its Super Agent platform. The company extended its Series B to $385 million and reported surpassing $200 million in annualized revenue by April 2026, at a valuation of approximately $1.6 billion, backed by Emergence Capital, Tencent, LG Technology Ventures, SBI, and Pavilion Capital. Research firm Sacra estimated approximately $100 million in annualized revenue as of January 2026, with more than 2 million monthly active users and roughly 100,000 paying seats at $30 per user per month. Its architectural choice — a mixture-of-agents orchestration layer routing tasks across more than 70 models from OpenAI, Anthropic, and others — is the most material commercial signal in this group: it monetizes through seats and usage rather than open-source stars or model-token distribution.

Manus is the cross-border-acquisition paradigm — and a case study in why that path may no longer be viable. Built by Butterfly Effect, founded in 2022 in Wuhan by Xiao Hong, Manus launched March 6, 2025 and went viral within 72 hours. It raised $75 million in April 2025 led by Benchmark, with backing from Tencent and HongShan. The company reported $125 million in annualized revenue by December 2025 — eight months after launch, a figure it claimed was the fastest $0-to-$125 million ramp on record. Meta reportedly agreed to acquire Manus in late December 2025 in a deal valued at more than $2 billion, with an explicit contractual condition of no continuing Chinese ownership. China's National Development and Reform Commission blocked the transaction on April 27, 2026, ordering the parties to unwind the deal on national-security grounds; according to multiple reports, the decision was elevated to Xi Jinping's National Security Commission. Two of Manus's co-founders were barred from leaving China during the review. Reuters reported in May 2026 that Manus's co-founder was attempting to raise approximately $1 billion to back out of the Meta acquisition.

What Is Driving the Agent Business Model Divergence?

These are not competing companies in the same market. OpenClaw and Hermes Agent grow through community adoption and carry no consumer revenue model. Genspark is the closest example in this group to a conventional subscription SaaS story — though Cursor, Codex, Claude Code, and Devin are stronger examples if the category is expanded to include coding agents. Manus reported more direct annualized revenue than the open-source pair combined, but the figure is company-reported and its exit path has become entangled in cross-border regulation that neither its founders nor its acquirer anticipated.

This is the reason calling any of these four the "winner" is premature: each is winning a different metric — stars, tokens, annual recurring revenue, acquisition interest — and none is yet winning on the metric the others care about.

Agent Layer Has No Dominant Business Model Yet

What the divergence reveals about the agent layer is the part worth retaining. The category is producing genuine product-market fit at multiple points on the value chain — local infrastructure, research-lab distribution, enterprise SaaS replacement, cross-border acquisition — but it has not produced a dominant business model, and the structural conditions for one are in conflict. Open-source distribution at scale is incompatible with $30-per-seat-per-month pricing. Cross-border acquisitions in this category have now become political events: Beijing blocked the Manus deal specifically because it viewed a Chinese-origin technology being acquired by an American platform as a national-security matter, not a commercial one. The parallel on the US side is the reported collapse of OpenAI's $3 billion Windsurf acquisition in mid-2025 after Microsoft reportedly blocked it due to exclusivity clauses, an indication that even within a single jurisdiction, the agent M&A market faces structural obstacles that pure commercial logic alone cannot clear.

Foundation labs that release free agents to drive model usage will eventually compete with the SaaS companies whose models they host — a competitive tension already visible in the fact that OpenAI funds OpenClaw while also selling Codex, which competes for the same developer workflows.

How Developers, Investors, and Builders Should Read the Map

For investors and builders, the harder implication is that the winner pattern that worked for prior software waves — identify the most-funded, fastest-growing, best-distributed company in a category and ride it — does not yet have a clean equivalent in agents. The largest commercial footprints are with coding incumbents already embedded inside developer workflows. The most interesting business-model experiments are at the edges: an Austrian solo developer's MIT-licensed project has more GitHub stars than any software repository in history; a three-month-old research lab agent has topped global inference rankings; a Palo Alto startup has crossed $200 million in annual run rate in under two years; and a Wuhan-founded agent that crossed $125 million in revenue found that its exit was a geopolitical event rather than a commercial transaction.

Which axis turns out to matter most — stars, tokens, seats, or M&A multiples — will determine which of the four paradigms above was right. None of them is right yet, and the incumbents are not standing still.

Frequently Asked Questions

What is the difference between open-source AI agents and subscription AI agents?

Open-source AI agents such as OpenClaw and Hermes Agent are freely downloadable, run on the user's own hardware, and have no subscription fee — their revenue model is indirect, typically driving adoption of the models or infrastructure their creators sell. Subscription AI agents such as Genspark charge per seat (around $30 per user per month as of mid-2026) and route tasks through a cloud orchestration layer managing multiple large language models on behalf of the user.

What happened to the Manus AI acquisition by Meta?

Meta agreed to acquire the AI agent startup Manus in a deal reported at more than $2 billion in late December 2025. China's National Development and Reform Commission blocked the transaction on April 27, 2026, ordering the two parties to unwind it on national-security grounds. The decision was elevated to Xi Jinping's National Security Commission. Two of Manus's co-founders were barred from leaving China during the regulatory review, and the company's co-founder was reported to be seeking approximately $1 billion in new funding to exit the deal.

What are the AI agent security risks for enterprise deployments?

Enterprise security teams have flagged autonomous agents as a privileged-access risk category: agents with broad tool access — email, file system, browser, code repositories — create an attack surface that traditional endpoint controls were not designed to handle. OpenClaw accumulated more than 138 documented security advisories between February and April 2026, including a CVSS 9.9-rated vulnerability that allowed full administrative access via a single API call. Cisco's AI Threat and Security Research team concluded that OpenClaw was "from a security perspective, an absolute nightmare," and Gartner recommended enterprises immediately rotate any corporate credentials an OpenClaw instance had accessed.

What are the leading AI coding agent platforms in 2026?

The largest coding-focused AI agent platforms by user count and revenue as of mid-2026 are GitHub Copilot (approximately 4.7 million paying users), Anthropic's Claude Code (approximately $1 billion in annualized revenue within six months of launch, with 46% "most loved" rating in a February 2026 Pragmatic Engineer survey of 15,000 developers), OpenAI Codex (more than 2 million weekly users, up five times in three months), and Cursor from Anysphere (valued at approximately $29.3 billion after a $2.3 billion raise in November 2025, with advanced funding talks reported at above $50 billion in April 2026).