Sixty thousand GitHub stars isn't usually cause for celebration in enterprise software. But when OpenHands, an agentic coding framework, hit that milestone as of November 21, 2025, something unusual happened. Developers kept coming. By March 2026, the count had climbed to 68,682—a momentum that suggests more than idle curiosity. It suggests belief.
This wasn't an isolated phenomenon. Dify reached 100,000 stars by mid-2025. ComfyUI, maintained by someone going only by "comfyanonymous," pulled in 89,200 by September. These numbers aren't vanity metrics, though they can be read that way. They're proxies for a deeper shift in how AI systems get assembled, who's doing the assembling, and what happens when developers decide they'd rather build the infrastructure themselves than rent it from the usual suspects.
The question facing the industry is no longer whether agentic AI will reshape enterprise software—that debate feels settled. The question is who ends up owning the orchestration layer when the dust clears.
When the Grassroots Move This Fast
Auto-GPT famously racked up 100,000 stars in under three months during spring 2023, setting what looked at the time like an unrepeatable benchmark. The current wave of frameworks feels different, though. More deliberate. Less proof-of-concept theater, more production evaluation. These tools are being kicked around by engineers who need to ship actual products.
The 2025 Stack Overflow Developer Survey offers a useful lens. Eighty-four percent of developers reported using or planning to use AI tools, up from 76% the year prior. Yet trust fell to 29%, down from 40%. That widening gap—between adoption and confidence—is where the orchestration wars are really being waged. Developers want the leverage that agentic AI promises. They're also acutely aware of the failure modes: prompt injection attacks, runaway costs, models that hallucinate at the worst possible moment.
Open-source frameworks have moved aggressively to fill that trust vacuum. LangChain closed a $125 million round at a $1.25 billion valuation in October 2025, positioning itself as a full-stack agent platform. LlamaIndex raised $19 million in a Series A back in March 2025, betting on knowledge agents that can handle messy enterprise data. CrewAI, with $18 million in announced funding from October 2024, built its entire pitch around multi-agent collaboration. Each is wagering that enterprises will choose composability and transparency over the convenience of vendor lock-in.
There's reason to think they might be right. Gartner projected in August 2025 that 40% of enterprise applications would embed task-specific AI agents by the end of 2026, up from less than 5% in 2025. By 2028, the firm estimated, at least 15% of day-to-day work decisions would be made autonomously by agentic systems—up from essentially zero in 2024. That's not iteration. That's a wholesale restructuring of work itself.
The Cloud Giants Aren't Idle
AWS brought multi-agent collaboration for Amazon Bedrock to general availability on March 10, 2025, complete with a Syngenta case study claiming a 5% crop yield improvement. In July, AWS introduced AgentCore for secure agent operations at scale. Swami Sivasubramanian, the company's VP overseeing AI, called agents "the most impactful change since the dawn of the internet" at an event in New York that month. Hyperbole, perhaps. Or perhaps not.
Google has matched the pace. Vertex AI Agent Builder received runtime billing updates in March 2025 and enhanced governance tooling by December. The company's marketing materials tout "hundreds of thousands of agents" deployed and "88% positive ROI among early adopters" in 2025—though those figures come from internal reports and warrant the usual skepticism. PwC and Google Cloud announced a portfolio of 120 AI agents in August, built on the Vertex stack.
Microsoft's path has been messier. AutoGen v0.4 launched in February 2025 with a complete redesign aimed at scale and extensibility. The company announced plans to converge AutoGen with Semantic Kernel into a unified agent framework, with a v1.0 release reportedly targeted for around April 2026. But as Forbes observed on April 6, 2026, Microsoft's agent stack has managed to confuse developers while rivals simplify—a damning critique in a space where developer experience is the entire game.
OpenAI made its own move with a March 11, 2025 launch of the Responses API and Agents SDK, bundling tools for web search, file handling, and computer use. The company announced plans to sunset its Assistants API after achieving feature parity, targeting the first half of 2026 for deprecation. AgentKit arrived in October 2025 with a visual builder and connector registry. By mid-April 2026, the SDK had evolved to include a model-native harness and sandboxed execution—each iteration tightening OpenAI's grip on the full stack.
The Quieter War Over Standards
Beneath the platform battles, a different kind of contest is unfolding around protocols. Anthropic open-sourced the Model Context Protocol on November 25, 2024, then donated it to the Linux Foundation's newly formed Agentic AI Foundation on December 10, 2025. By that point, more than 10,000 MCP servers were reportedly in operation, with adopters spanning Claude, Microsoft Copilot, Gemini, ChatGPT, and Cursor.
The protocol's pitch is straightforward: decouple tool and data connectivity from the orchestration layer itself. ExpressVPN's launch of an MCP server beta on March 5, 2026 showed the concept in action—giving AI agents programmatic control over VPN clients through a standardized interface. That's the kind of interoperability that could defang vendor lock-in, if it takes hold.
OpenAI made a parallel move by co-founding the Agentic AI Foundation in December 2025 and donating its AGENTS.md standard, which the company said had seen over 60,000 open-source adoptions since August 2025. Whether these standardization efforts represent genuine neutrality or strategic maneuvering is an open question. But the fact that Anthropic, OpenAI, and the Linux Foundation are all racing to define the connective tissue of agentic systems suggests everyone understands what's at stake.
Standardization carries its own risks, though. A security analysis published in late January 2026 uncovered vulnerabilities in Anthropic's official Git MCP server—a reminder that moving fast on standards can create systemic exposure. The incident underscores a broader challenge: as agents gain autonomy and tool access, the attack surface expands dramatically.
Security Has Become Non-Negotiable
Research published through 2025 and early 2026 has cataloged the escalating threats. Studies have benchmarked web-agent hijacking scenarios and proposed multi-agent defense architectures. OpenAI blog posts from February and March 2026 detailed ongoing hardening efforts for ChatGPT Atlas, its agent-in-browser product, against prompt injection. These aren't theoretical concerns. They're active threat vectors.
NVIDIA has positioned NeMo Guardrails and NIM microservices as the answer, emphasizing programmable guardrails and observability for production agents. Cisco announced an integration with NeMo Guardrails on October 28, 2025, extending the framework's reach into enterprise security stacks. The message is unambiguous: if you're deploying agents in production, you need guard rails baked in, not bolted on.
The regulatory environment is tightening in parallel. The EU AI Act entered force in August 2024, with obligations for general-purpose AI models kicking in from August 2025, and high-risk system requirements arriving in August 2026. In the U.S., OMB memos M-25-21 and M-25-22, issued in April 2025, formalized governance and procurement guardrails that flow down to vendors and integrators. NIST's AI Risk Management Framework for Generative AI, published in April 2024, is being invoked to justify agent governance requirements across federal agencies and beyond.
The most telling shift may be cultural. Developer trust in AI tools dropped 11 percentage points in a single year, according to the 2025 Stack Overflow survey. That erosion of confidence is creating demand for evaluation pipelines, observability platforms, and identity-aware permission systems at the orchestration layer—capabilities that open-source frameworks and cloud providers alike are scrambling to deliver.
Follow the Capital
The venture community has taken notice. LangChain's $1.25 billion valuation in October 2025 marked a milestone: a pure-play orchestration company hitting unicorn status before the agent market had fully matured. LlamaIndex's $19 million Series A in March 2025, led by Norwest with participation from Greylock, validated the bet that retrieval-augmented generation and knowledge agents would form critical infrastructure. CrewAI's $18 million in funding, announced in October 2024 by Insight Partners, wagered on multi-agent collaboration as the dominant pattern.
Consolidation has already begun. ClickHouse's acquisition of Langfuse in January 2026 sent a clear signal that the observability and evaluation layer is ripe for M&A. Langfuse had previously open-sourced its full product in June 2025 and reported over 8,000 monthly active self-hosted instances. Its absorption into ClickHouse suggests that observability and analytics will increasingly be bundled into data infrastructure rather than sold as standalone tools.
Enterprise case studies are starting to land. Syngenta's Cropwise AI, built on AWS Bedrock's multi-agent collaboration and detailed in a March 2025 blog post, reported a 5% increase in crop yields—tangible returns from a multi-agent architecture. PwC's portfolio of 120 AI agents, announced in August 2025, demonstrates how professional services firms are packaging agents as repeatable solutions across client engagements.
The market remains bifurcated. Well-funded open-source frameworks like LangChain, LlamaIndex, and CrewAI are building ecosystems around community-driven development and flexible deployment. On the other side, AWS, Google, Microsoft, and OpenAI are racing to provide fully managed orchestration, governance, and observability—betting that enterprises will pay a premium to avoid infrastructure headaches.
What Comes Next
The next 12 to 18 months will likely determine whether the orchestration layer commoditizes or consolidates. Standardization efforts like MCP and AGENTS.md could reduce switching costs and fragment the market. Or they could simply raise the table stakes, pushing competition upward into deployment automation, security tooling, and enterprise data integration—areas where the cloud giants have structural advantages.
Gartner's forecast that agentic AI could generate roughly 30% of enterprise application software revenue by 2035—approximately $450 billion, up from around 2% currently—suggests the prize is large enough for multiple winners. But the firm's 2026 target of 40% of enterprise apps featuring task-specific agents implies a brutal near-term race. Enterprises won't adopt 15 different orchestration frameworks. They'll standardize on two or three. The rest will fade into the long tail of GitHub history.
Security and governance will likely separate winners from also-rans. The frameworks and platforms that make it easiest to implement identity-aware permissions, audit trails, cost controls, and prompt injection defenses will win enterprise mindshare. The ones that treat security as an afterthought will get benched, regardless of API elegance.
What's certain: the orchestration wars are no longer a sideshow. They're the main event. The stack is being rewritten from the ground up, with agents as the new primitives. Whether that stack will be owned by pseudonymous open-source developers, billion-dollar platforms, or some hybrid coalition remains unsettled.
But the fact that it's even a contest—that grassroots frameworks can go viral faster than tech giants can ship GA products—reveals where the power currently sits. It sits with the developers who build, evaluate, and deploy these systems. Right now, they're watching closely to see who earns their trust. And trust, in this market, isn't handed out lightly.
