Four AI Agent Projects, Four Incompatible Business Models, and No Winner

A Tech Times analysis published May 24 maps the four AI agent projects generating the most coverage in mid-2026 and arrives at an uncomfortable conclusion: they are not competing with each other. OpenClaw, Hermes Agent, Genspark, and Manus each represent a fundamentally different theory of what an agent business looks like. None has proven the others wrong, and the structural conditions for a single dominant model are in active conflict.

The Four Paradigms

OpenClaw: open-source as infrastructure. Peter Steinberger’s MIT-licensed personal agent hit 370,000 GitHub stars by late May 2026, runs locally on user hardware, and connects to more than 50 messaging platforms. It has no consumer revenue model. Its most revealing public data point: Steinberger ran 100 agents continuously for 30 days and generated a $1.3 million OpenAI token bill, covered by OpenAI itself. Steinberger joined OpenAI in February 2026 to lead personal agents research. The project transferred to an independent foundation with OpenAI as financial sponsor.

Hermes Agent: research lab as distributor. Released February 25, 2026 by Nous Research, Hermes Agent crossed 140,000 GitHub stars in under three months and overtook OpenClaw on May 10 to claim the top spot on OpenRouter’s daily inference rankings, processing 224 billion tokens in a single day against OpenClaw’s 186 billion. NVIDIA positioned the agent as a showcase workload for its RTX and DGX Spark hardware. There is no Hermes Agent subscription. The project exists to drive usage of Nous’s models and inference infrastructure.

Genspark: subscription SaaS. The Palo Alto-based super-agent platform, founded by former Baidu VP Eric Jing, extended its Series B to $385 million and surpassed $200 million in annualized revenue by April 2026, at a $1.6 billion valuation. Research firm Sacra estimated roughly $100 million ARR as of January 2026, with approximately 100,000 paying seats at $30 per user per month. Genspark routes tasks across more than 70 models through a mixture-of-agents orchestration layer.

Manus: cross-border acquisition. The Wuhan-founded autonomous agent hit $125 million in annualized revenue eight months after its March 2025 launch, then agreed to a $2 billion-plus Meta acquisition in December 2025. China’s NDRC blocked the transaction in April 2026 on national-security grounds. Two co-founders were barred from leaving China during the review.

The Elephant in the Room: Coding Agents

The Tech Times piece makes a point that most agent coverage ignores: the dominant commercial players in the agent layer are coding-tool incumbents. OpenAI’s Codex serves more than 2 million weekly users. Anthropic’s Claude Code reached roughly $1 billion in annualized revenue within six months. Cursor raised $2.3 billion at $29.3 billion in November 2025 and is reportedly in talks for a round above $50 billion. GitHub Copilot has roughly 4.7 million paying users.

Compared to those numbers, the four paradigm projects are small. What they offer instead is strategic clarity about which business model each is testing.

Why This Cannot Resolve Quickly

The structural conflicts between these paradigms are not minor. Open-source distribution at massive scale is incompatible with $30-per-seat pricing. Cross-border acquisitions in the agent category have become geopolitical events, not commercial ones. And foundation labs that release free agents to drive model usage will inevitably compete with the SaaS companies whose models they host.

That last tension is already visible: OpenAI funds OpenClaw while selling Codex, and both target developer workflows.

The Builder’s Dilemma

For anyone building on the agent layer, the practical question is which paradigm your infrastructure depends on. If you build on OpenClaw, you depend on a foundation funded by a company that also sells a competing product. If you build on Hermes, you depend on a research lab whose business model is driving inference volume, not supporting your production workload. If you pay $30/seat for Genspark, you depend on a mixture-of-agents layer that routes across 70+ models from providers who are also building their own agents.

The Tech Times analysis concludes that the winner-take-all pattern from prior software cycles does not map cleanly onto the agent layer. Each paradigm is winning on a different metric: stars, tokens, revenue, acquisition interest. None is winning on the metric the others care about.

That fragmentation is the market signal. Not which project is ahead, but that the question “ahead on what?” still does not have a consensus answer.