How Harness-as-a-Service Will Change Agents

April 30, 2026

28 min episode · 2 min read

Episode

28 min

Read time

2 min

AI-Generated Summary

Published Apr 30, 2026

Key Takeaways

✓Big Tech AI Revenue: Google Cloud grew 63% year-over-year with a $460B order backlog, AWS grew 28% reaching $152B ARR, Azure grew 39%, and Meta posted 33% revenue growth. These numbers signal AI demand is no longer speculative — it is measurable, accelerating, and supply-constrained across every major hyperscaler.
✓Harness Engineering Phase: Agent capability now evolves along two vectors: model improvements and harness improvements. The same model running inside different harnesses produces measurably different results — GPT-5.5 jumped from 61.5% to 87.2% on functionality benchmarks simply by switching from its native Codex harness to Cursor's harness, per Endor Labs testing.
✓Harness as a Service Definition: Treat agent runtimes like infrastructure primitives. Cursor SDK, OpenAI Agents SDK, Anthropic Managed Agents, and Microsoft Foundry all pre-build the agent loop, tool dispatch, sandboxing, error handling, and context compression — developers supply only the model choice, tool access, and task definition, reducing assembly work dramatically.
✓Non-Developer Builder Opportunity: The Cursor SDK expands the builder audience beyond traditional developers. Non-technical builders can drop the SDK's GitHub cookbook into Claude or ChatGPT with project context and generate viable agent architectures. Agents handling the coding layer means the barrier to building harness-powered products has dropped to task definition and tool selection.
✓Search Cannibalization Thesis Reversed: Google's search ad revenue grew 19% year-over-year with queries hitting all-time highs despite widespread adoption of AI chatbots. The predicted substitution effect — users abandoning Google for LLM-based answers — has not materialized, suggesting search and conversational AI are currently complementary rather than competitive behaviors.

What It Covers

Big tech Q1 AI earnings reveal accelerating cloud growth across Google, Microsoft, Amazon, and Meta, while Cursor's new SDK exemplifies a broader infrastructure shift called "Harness as a Service" — a category where companies sell pre-built agent runtimes the same way AWS sells compute.

Key Questions Answered

•Big Tech AI Revenue: Google Cloud grew 63% year-over-year with a $460B order backlog, AWS grew 28% reaching $152B ARR, Azure grew 39%, and Meta posted 33% revenue growth. These numbers signal AI demand is no longer speculative — it is measurable, accelerating, and supply-constrained across every major hyperscaler.
•Harness Engineering Phase: Agent capability now evolves along two vectors: model improvements and harness improvements. The same model running inside different harnesses produces measurably different results — GPT-5.5 jumped from 61.5% to 87.2% on functionality benchmarks simply by switching from its native Codex harness to Cursor's harness, per Endor Labs testing.
•Harness as a Service Definition: Treat agent runtimes like infrastructure primitives. Cursor SDK, OpenAI Agents SDK, Anthropic Managed Agents, and Microsoft Foundry all pre-build the agent loop, tool dispatch, sandboxing, error handling, and context compression — developers supply only the model choice, tool access, and task definition, reducing assembly work dramatically.
•Non-Developer Builder Opportunity: The Cursor SDK expands the builder audience beyond traditional developers. Non-technical builders can drop the SDK's GitHub cookbook into Claude or ChatGPT with project context and generate viable agent architectures. Agents handling the coding layer means the barrier to building harness-powered products has dropped to task definition and tool selection.
•Search Cannibalization Thesis Reversed: Google's search ad revenue grew 19% year-over-year with queries hitting all-time highs despite widespread adoption of AI chatbots. The predicted substitution effect — users abandoning Google for LLM-based answers — has not materialized, suggesting search and conversational AI are currently complementary rather than competitive behaviors.

Notable Moment

Sam Altman told interviewer Ben Thompson that the harness surrounding a model is nearly impossible to separate from the model itself — when an agent completes a task inside Codex, he genuinely cannot determine whether the model or the runtime environment deserves the credit for the result.

Know someone who'd find this useful?