The AI Agent Harness Is the Product: Four Vendors Converge, Then Diverge on Price

Four of the largest AI platform vendors now agree on something that wasn’t obvious a year ago: the model itself is becoming a commodity, and the infrastructure surrounding it — the harness — is where the strategic value lives. The disagreement, and it’s a significant one, is over who builds, owns, and pays for that harness.

Context

The AI agent harness is everything that isn’t the model. The New Stack’s analysis defines it as “model invocation and context management, tool orchestration, sandboxed execution, persistent session and execution state, scoped permissions, error recovery, observability, and tracing.” If that list sounds like production infrastructure, it should — it’s analogous to the containerization and orchestration layer that Kubernetes eventually owned for application deployment. The vendors betting on controlling this layer are not subtle about their intent.

The timing reflects where enterprise AI deployments are in their maturity curve. Proof-of-concept phases run without harnesses; production deployments require them. As agent workflows expand into regulated and high-stakes environments — finance, healthcare, legal — the requirements for persistent state, scoped permissions, and auditability push enterprises to evaluate their harness options seriously. Platform teams that bought time on the POC phase are now facing real infrastructure decisions.

The Detail

Anthropic, OpenAI, Google, and Microsoft all agree the harness layer is strategically important and want to own it. Where they split is on the monetization model.

Anthropic’s approach is fully managed: Managed Agents bundles the harness — execution environment, state management, session persistence — into a single offering at eight cents per session hour plus standard token rates. Everything runs on Anthropic’s infrastructure; the developer doesn’t manage the runtime. This mirrors how AWS Fargate operates relative to raw EC2 — higher abstraction, higher unit cost, lower operational burden.

OpenAI took the opposite position. Its Agents SDK ships as free, open-source software with no separate harness fees and no session-hour metering. Developers bring their own execution infrastructure through providers like E2B, Modal, and Vercel, and pay only for those services alongside token consumption. The model is control in exchange for operational responsibility — similar to self-managed Kubernetes versus a managed cluster.

Google and Microsoft sit between these positions, using component-based metering where sessions, memory, code execution, and observability are billed separately within their broader platform ecosystems. More granular, more complex to model financially, but better suited to the existing procurement and chargeback patterns enterprises use for cloud spend.

As AWS’s own MCP investments signal, the cloud providers are simultaneously building managed harness capabilities while standardizing the protocol layer underneath — a pattern that historically concentrates value at the managed-service tier over time.

Source: The New Stack

Content created with AI assistance and reviewed for accuracy.