Coding agents need sandboxed filesystems that persist across hundreds of LLM calls. Browser agents need full DOM access, cookie persistence, and bot detection bypass. Research agents run for 30 minutes or more. None of this fits into AWS Lambda’s 15-minute timeout, stateless execution model, or cold-start latency profile. A new class of infrastructure companies is filling that gap, and by mid-2026 the category has enough revenue, funding, and customer adoption to qualify as a distinct infrastructure layer.

What Cloud Providers Did Not Build

The mismatch between agent workloads and existing cloud primitives is specific and measurable. Lambda caps execution at 15 minutes. A coding agent debugging a complex test failure can run for 30 minutes of wall-clock time. A research agent compiling competitive analysis across dozens of sources can run for an hour. Lambda’s stateless model breaks further: every LLM call in an agent loop depends on cumulative state (files edited, pages navigated, shell commands executed), and storing that in external databases introduces failure modes that negate the serverless premise.

EC2 and Kubernetes solve the runtime issues but not the workload-specific ones, according to an analysis by elongated_musk on Medium. An agent needs a real browser with JavaScript execution and cookie persistence. It needs code execution sandboxes that run untrusted Python or JavaScript safely. It needs per-task cost attribution across LLM tokens, compute seconds, browser sessions, and external API calls. These requirements are sufficiently different from request-response web traffic that a new infrastructure category has formed around them.

The Companies Filling the Gap

Three platforms have captured defensible positions in agent runtime. E2B raised $21M in a Series A led by Insight Partners in July 2025, bringing total funding to $32M. The company reports that 88% of Fortune 100 companies have signed up on its platform, with hundreds of millions of cloud sandboxes launched. Its customer list includes Perplexity (which shipped advanced data analysis in one week using E2B sandboxes) and Manus, which uses E2B to give agents virtual computers.

Browserbase provides managed browser infrastructure for agents that need to navigate the web. The company is hosting Navigate 2026, its first dedicated conference, in a signal that the customer base has grown large enough to sustain an event. Modal offers sandboxed compute, inference, and training infrastructure with products explicitly designed for AI workloads.

The category is not theoretical. These companies have paying enterprise customers, growing revenue, and the kind of late-stage venture valuations that suggest IPO timelines within two to three years.

MCP as the Interface Standard

The Model Context Protocol has emerged as the connective layer between agents and tools. MCP servers now exist for GitHub, Slack, Linear, Notion, Postgres, Stripe, and dozens of other services. The protocol standardizes how agents discover and invoke external tools, reducing integration friction for builders.

This standardization has a double-edged economic effect. Lower switching costs expand the addressable market for runtime platforms, because any agent framework can plug into any MCP-compatible tool. But those same low switching costs compress margins, because customers can migrate between platforms more easily. The dynamics mirror what happened with containerization: Docker and Kubernetes standardized deployment, which made the overall market far larger while simultaneously making any single vendor more replaceable.

Microsoft accelerated this convergence last week when Fabric introduced MCP servers that let agents operate natively on Power BI resources, data workspaces, and pipelines. When the largest enterprise software company builds native MCP support into its data platform, the protocol’s status as a settled standard is no longer debatable.

The Containerization Parallel

The closest historical analog is the rise of container infrastructure in 2013 to 2017. Docker launched in 2013 as a packaging format for applications. Kubernetes followed as an orchestration layer. Within four years, every major cloud provider offered managed container services, and a generation of infrastructure companies (HashiCorp, Datadog, Confluent) built billion-dollar businesses on the tooling layer around containers.

Agent runtime is following the same trajectory, compressed into a shorter timeframe because the underlying AI models are improving faster than containers did. The core pattern is identical: a new workload type that existing cloud primitives handle poorly generates demand for specialist infrastructure, which standardizes around protocols that enable an ecosystem of adjacent tools and services.

The $450 Billion Question

The viability of this category depends on agents actually reaching production at scale. If reliability issues or regulatory friction keep agents stuck at demo stage, customer demand stalls and the runtime platforms’ growth curves flatten. The BIS warning about circular financing in AI infrastructure applies here too: the runtime platforms are funded by VCs betting on agent adoption, which is funded by enterprises betting on productivity gains that have not yet been proven at scale.

But the adoption signals are accumulating. ServiceNow’s $2.85 billion acquisition of Moveworks earlier this year, reported by IT Brief Australia, validated that enterprise buyers consider agent-first architectures worth paying for. E2B’s Fortune 100 penetration suggests that procurement teams at the largest companies have already approved agent infrastructure spending.

The Investment Thesis

Agent runtime is not a feature. It is infrastructure, the same way load balancers, CDNs, and container orchestrators are infrastructure. The platforms that win this category will look less like application companies and more like the next Datadog or Cloudflare: horizontal infrastructure layers that every AI-powered product depends on, priced on usage, sticky through integration depth.

The risk is that hyperscalers build competitive offerings before the startups reach escape velocity. AWS, Google Cloud, and Azure all have the compute resources and enterprise relationships to offer agent runtime as a managed service. The question is whether they move fast enough, or whether Browserbase, E2B, and Modal entrench themselves the way Snowflake entrenched itself against BigQuery and Redshift.

For builders, the practical takeaway is simpler: the agent execution layer exists, it works, and it is funded well enough to rely on. The days of stitching together Lambda functions and EC2 instances to run agent workloads are ending.