OpenAI Designed Agents to Cost You More. That's Not a Bug.
OpenAI's agent architecture isn't a capability bet — it's a revenue extraction mechanism, and the enterprises signing multi-year agentic contracts are only now starting to understand they're the product, not the customer.
The skills-vs-agents debate gets framed as a technical question. It isn't. It's a question of whose incentives are embedded in the architecture.
OpenAI's Agents SDK rebuilds API interactions around loops. Each loop is multiple inference calls. Multiple calls is more tokens. More tokens is more revenue. This isn't incidental — it's structural. The orchestration layer isn't a developer convenience; it's a billing mechanism wearing a capability argument.
Anthropic's Model Context Protocol takes a different position. MCP's core framing — standardize M tools and N models instead of M×N integrations — is a standardization argument, not a capability argument. Standardize the interface, commoditize the tool layer, make the model the scarce resource. It doesn't scale Anthropic's revenue with your token consumption the way agent loops do. That's a different lock-in mechanism, but it's still lock-in.
The production data makes the stakes concrete. 92.5% of deployed "agents" deliver output to humans rather than downstream systems. 68% require human intervention within 10 steps. What enterprises have actually shipped isn't autonomous action — it's a well-funded draft review workflow paying agent-tier infrastructure costs of $3,000–$13,000 per month for output that skill-based pipelines generate cheaper. BCG X CTO Matt Kropp said it plainly: "most deliver very little ROI." This isn't a fringe view — it's the private consensus of people running these deployments.
The complication worth taking seriously: agent loops do unlock things skills cannot. Multi-step reasoning, genuine error recovery, novel task decomposition — these capabilities are real. The question is whether enterprises signing multi-year agentic contracts are buying those capabilities or buying the marketing category around them.
Gartner predicts 40% of agentic projects canceled by 2027 — not failed, canceled. The distinction matters because cancellations get repackaged. They'll read as "we matured our AI strategy," not "we overpaid for API loops for two years." The companies coming out ahead already accepted the 92.5% human-review figure as destination rather than waypoint, stripped the orchestration layers, and built skill pipelines that audit cleanly and run cheaply. They'll look prescient in 2027. But the harder question underneath all of this doesn't go away: if agent value collapses to faster human-assisted drafting, and simpler skills already deliver that at lower cost, what exactly is the agent platform selling — and to whom?