Gartner expects more than four in ten agentic AI projects to be cancelled by 2027, inside a market compounding at close to 50 percent a year. Almost none of those failures will be failures of the models. They happen further down, at the point where software meets permission, and a Bengaluru operator has now been the first commercial hire at two companies working that seam.
In June 2025, Gartner forecast that more than 40 percent of agentic AI projects would be cancelled by the end of 2027, attributing the failures to escalating costs, unclear business value and inadequate risk controls. In the same assessment the firm estimated that of the thousands of vendors then marketing agentic capabilities, only around 130 were building anything that met the definition. The finding was carried by Reuters and read across the industry as a verdict on the technology.
It was something narrower than that. What the forecast described was a market in which the binding constraint had stopped being capability and became deployment. Models could already reason, plan and execute. What they could not reliably do was reach the systems, the permissions and the people that turn an agent into working infrastructure inside an organisation. The gap that opened between those two facts is where a specific and largely uncredited kind of operator became valuable.
Karthikeya Meesala, currently Head of Growth at Oasis and previously among the earliest members of the founding team at Composio, has built his career on that side of the problem. His trajectory, from a pre-product agent infrastructure company in Bengaluru through a $25 million Series A to a four-person team building the control layer above it, tracks the moment the agentic AI industry stopped competing on what its software could do and started competing on whether anyone would use it.
A market that outgrew its own adoption curve
The headline numbers are not in dispute. Grand View Research valued the global AI agents market at $7.6 billion in 2025 and $10.9 billion in 2026, projecting $182.9 billion by 2033 at a compound annual rate of 49.6 percent. MarketsandMarkets is more conservative, putting the market at $7.84 billion in 2025 and $52.62 billion by 2030. Estimates diverge by an order of magnitude at the far end because the category boundary itself is unsettled, but every major forecaster describes the same curve.
The number underneath the curve is less flattering. Gartner’s 2026 Hype Cycle for Agentic AI places the category at the peak of inflated expectations and reports that only 17 percent of organisations have deployed AI agents, while more than 60 percent expect to do so within two years. That spread, between intent and installation, is the whole commercial problem of the sector. It is also why the companies that solved connection and control ahead of the demand curve became disproportionately important.
Composio: building into the constraint
Composio, incorporated as Sampark Inc., was founded in 2023 by Soham Ganatra and Karan Vaidya, IIT Bombay alumni who first met at a Physics Olympiad camp. The company built infrastructure that lets AI agents connect to the software organisations already run, spanning more than 10,000 external tools, with authentication, permissioning and error handling managed at the platform layer rather than rebuilt for every integration. It raised a $4 million seed round co-led by Together Fund and Elevation Capital, followed by a $25 million Series A led by Lightspeed Venture Partnersannounced in July 2025, with participation from Guillermo Rauch of Vercel, Dharmesh Shah of HubSpot, Gokul Rajaram and Soham Mazumdar of Rubrik. Total funding stands at roughly $29 million.
Meesala joined pre-product, before the platform had paying customers, and stayed through the Series A. By January 2026, Mr. Ganatra told Forbes India that the platform had passed 200,000 developers and was processing more than 100 million tool calls a month. He was named to the publication’s 30 Under 30 list for 2026 in the same cycle.
In January 2025, with the platform still pre-revenue, Meesala built and ran Composio’s startups programme: free access to the paid tier for early-stage teams, alongside direct working time with the founders and the engineering team. Three hundred startups came through it, and that cohort became both the company’s first revenue base and the customer evidence underneath the July Series A. Eighteen months on, the programme is still running.
What a protocol did to the moat
Something structural happened to that business in the interim. In November 2024, Anthropic published the Model Context Protocol, an open standard for how agents discover and call external tools. Adoption was unusually fast for infrastructure. By Anthropic’s December 2025 ecosystem update, more than 10,000 active public MCP servers were in circulation and the official Python and TypeScript SDKs were drawing over 97 million downloads a month, with support shipped across ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code. Stacklok’s 2026 software report found 41 percent of surveyed software organisations running MCP servers in limited or broad production.
Standardisation is good for an industry and hard on the companies that were previously differentiated by integration coverage. Once connecting an agent to Slack or GitHub became a solved problem with a published specification, breadth of integrations stopped being a durable advantage. What remained was distribution: developers knew a platform existed, tried it during an evaluation window measured in hours, and left it running in production six months later. In developer infrastructure that work is not marketing in the conventional sense. It is closer to the product, and it is the function Meesala was hired into before the company had a product to distribute.
Oasis: the same problem, one layer up
Meesala’s current role is Head of Growth at Oasis, operating as Mercury Intelligence Inc., where he is the second hire at a four-person company. The product is a no-code workspace for spinning up coordinated teams of AI agents in about thirty seconds, mixing agents from Anthropic’s Claude, OpenAI and other frameworks, with policies, approval gates and hand-off to humans over Slack and iMessage. The platform opened to the public in July 2026, and a second and larger release is scheduled for October.
The design reads as a direct answer to Gartner’s diagnosis. Of the three causes the firm gave for the projected cancellations, inadequate risk controls is the one Oasis addresses at the architectural level, through explicit policy and approval layers rather than through documentation. The second answer is subtler and concerns who is allowed to build. Composio sold to engineers. Oasis sells to operators who do not write code, which changes the adoption problem from developer trial to organisational trust, and moves the burden of proof from whether the software works to whether a manager will let it act.
The layer that gets measured last
Meesala’s work now spans both sides of the same bottleneck. At Composio, he built demand for the infrastructure that lets agents reach the tools an organisation already uses. At Oasis, he is working on the layer where non-technical operators decide what those agents are permitted to do, and whether the result is worth keeping.
The sector has spent three years pricing model capability and comparatively little pricing adoption, which is the variable most of the cancelled projects will turn out to have failed on. If Gartner’s forecast holds, the agentic AI companies still operating in 2028 will not be the ones with the largest models. They will be the ones that got used. That is a go-to-market problem before it is a technical one, and it is beginning to be treated as a discipline in its own right rather than as the work that happens after the engineering is done.