The API tax: Why AI agents stall without infrastructure context

The infrastructure gap facing AI agents

AI agents are increasingly being deployed to accelerate software development, but they frequently stall when they encounter APIs that lack sufficient infrastructure context. The problem, described as "the API tax," highlights a growing gap between what AI agents need to function effectively and what current API infrastructure provides.

According to reporting from The New Stack, AI agents speed up development workflows, but data access remains a significant slowdown. Articles on the publication have noted that while agents can write code, they often lack the contextual information about infrastructure, deployment environments, and system architecture needed to operate autonomously.

Agent-friendly APIs and the MCP question

The Model Context Protocol (MCP) has emerged as a way to get AI agents into APIs, but as one recent article on The New Stack points out, MCP does not decide what agents should see. This distinction is critical: connecting agents to APIs is only the first step. Controlling and contextualizing the data they access remains an unsolved challenge.

Building scalable, agent-friendly APIs for AI applications has become a priority for platform engineers. The publication has covered how organizations are rethinking API design to provide agents with the structured context they need, rather than exposing them to raw, unstructured data that can lead to errors and inefficiencies.

Latency, verification, and the runtime problem

Agentic AI has a latency problem that more compute will not solve, according to analysis on The New Stack. The issue is not raw processing speed but the time agents spend waiting for data, verifying outputs, and navigating complex infrastructure contexts. For cloud-native software, verification is fundamentally a runtime problem, requiring agents to test their code against real environments rather than static configurations.

Tools like Greptile, Cursor, and Devin agree that agents should run their code, but what they run it against matters significantly. The infrastructure context provided during execution determines whether agents can verify their work effectively or simply produce plausible but unverified output.

Security, traces, and the path forward

As AI agents become more autonomous, security concerns have intensified. WebAssembly has been proposed as a potential solution to AI agents' most dangerous security gap, providing sandboxed execution environments that limit the blast radius of agent actions. Meanwhile, when AI agent traces become application data, organizations must grapple with how to store, manage, and govern the operational records that agents generate.

The broader ecosystem is responding. AWS has open-sourced an AI agent it says is 45% cheaper than Claude Code and Codex. Google wants to make the web agent-ready. Cloudflare is building what it calls the economic layer of the AI web. These efforts suggest that the industry recognizes the API tax as a real barrier to agent adoption, and that infrastructure context is the key to unlocking agent potential.

Sources: The New Stack