51% of web traffic is now AI, but most APIs still return HTML (medium.com)

Web traffic has officially tipped into machine-first territory — Imperva reports 51% of traffic is non-human (Gartner expects 80% by 2027) — and the race is now for agent infrastructure rather than just bigger LLMs. Major platforms are already shipping agent-ready primitives: Anthropic’s MCP (Model Context Protocol), OpenAI’s Agents SDK, Google Vertex AI Agent Builder (MCP-compatible) and Microsoft’s Azure AI Foundry (100k+ production agents). The business case is simple: agent-friendly APIs that return structured machine-readable JSON instead of HTML drastically reduce token counts and costs (example: ~150 tokens/$0.05 vs ~15 tokens/$0.0025), raise agent success rates (62% → 98%) and yield 95%+ per-request savings — at scale this becomes millions in compute and new microtransaction revenue streams (x402 example: 1M daily calls = $1,000/day). Technically, winners are building a three-layer stack: Discovery (OpenAPI with x-agent-friendly metadata), Orchestration (Arazzo workflows) and Execution (MCP servers with large context limits — e.g., 128k tokens — and tool sets). New KPIs (AIX: discovery/completion/tokenEfficiency/recovery; TX: cost per successful agent interaction; MPX: micro-payment velocity) guide product and investment decisions; top teams hit <$0.001 per agent transaction. Complementary advances like GEPA (Genetic-Pareto Prompt Evolution) promise ~20% better prompts at 35x lower optimization cost. The implication: APIs that go “bilingual” become distribution, pricing and optimization moats — early movers capture multiplicative revenue and durable network effects.