Sunil Venkataram
Back to garden
•
🌰 seedling
Harness as a Service - Infrastructure for the Agent Era
type/noteaiagentsinfrastructuredeveloper-tools

Harness as a Service — Infrastructure for the Agent Era

The Category

In spring 2026, several major platforms launched HaaS products nearly simultaneously:

PlatformProductKey Feature
CursorCursor SDKExposes Cursor’s coding agent runtime (repo context, edit, search, terminal, streaming) for embedding in any app
AnthropicClaude Managed AgentsCloud-hosted meta-harness with brain/hands architecture
MicrosoftHosted Agents in FoundryDedicated enterprise-grade sandbox per agent with durable state, identity, governance
OpenAIAgents SDK updateUpdated agent runtime and tool dispatch

Satya Nadella’s framing: “Every agent will need its own computer.” The HaaS provider supplies that computer.

The Hobbyist-to-PC Analogy

The shift from open-source harness frameworks (like OpenClaw) to HaaS mirrors the shift from kit computers to the Apple II.

Kit era (OpenClaw, ~early 2026): The builder assembles everything — model selection, system prompt, tool definitions, agent loop, context management, error handling, state persistence, deployment, monitoring. Powerful but limited to a small audience willing to solder the components together.

PC era (HaaS, ~mid 2026): The agent loop is pre-built, tool dispatch is pre-built, sandboxing is pre-built. The productivity revolution of the 1990s happened because users got Dell desktops, not because more people learned to assemble motherboards. HaaS makes the same bet: democratize the runtime, expand the builder base.

This is a shift in kind, not just scale. The kit era didn’t disappear — OpenClaw and its descendants still serve builders who want full control — but HaaS expands the addressable market for agent-powered applications by orders of magnitude.

Harness Dominance Over Model Quality

Benchmark evidence from Endor Labs (April 2026) demonstrates that the harness matters as much as the model:

ModelHarnessFunctionality ScoreSecurity Score
Opus 4-7Claude Code87.2%—
Opus 4-7Cursor91.1%22.9%
GPT-5.5Codex61.5%—
GPT-5.5Cursor87.2%23.5%

GPT-5.5 jumped 25.7 percentage points on functionality by switching from Codex to Cursor’s harness. Same model, same week, two harnesses, two different results. This makes the harness a first-class product decision, not an implementation detail.

The Expanded Builder Class

A subtle consequence of HaaS: because agents now handle the coding, the audience who can build with an SDK like Cursor’s extends far beyond traditional developers. Non-developers who understand their domain can specify what they want, and the agent runtime handles the implementation within the HaaS platform.

Early Cursor SDK examples illustrate this:

  • Gmail integration — a Cursor agent embedded in Gmail reads email threads, edits code, and streams results back into the inbox
  • Visual bug-catching — an agent that watches the running app in its own browser, closing the feedback loop between code generation and visual verification
  • IT triage Chrome plugin — non-technical users dump browser code into tickets instead of describing bugs in prose

Key Takeaways

  1. HaaS is the natural next step after Agent Harnesses proved that the harness matters more than the model. If the harness is the moat, selling access to it is the business model.
  2. The hobbyist-to-PC transition suggests HaaS will expand the agent builder population far beyond its current base, just as PCs expanded the computer user population.
  3. Benchmark evidence makes harness selection a first-class engineering decision — choosing a HaaS provider may matter more than choosing a model provider.
  4. The convergence of multiple HaaS launches in spring 2026 suggests this is a platform shift, not a product feature.
Backlinks (1)
Connected Notes