The Great Compression: Model Providers Are Swallowing the Agent Harness Layer

OpenAI’s March 2026 acquisition of Promptfoo — the AI security testing platform used to evaluate prompt injection resistance and LLM compliance — marks the eighth harness-layer company the company has absorbed since June 2024. It follows Rockset (real-time retrieval), Context.ai (evaluation talent), Statsig ($1.1 billion, product analytics), and Neptune (ML training observability). Each acquisition represents a middleware function that will no longer be available as an independent vendor to enterprises building multi-provider architectures. The pattern is not incidental. It is the endgame of a vertical integration strategy that all six major model providers are executing simultaneously.

The agent harness layer has never been a stable category. It emerged as a practical necessity in 2023 and 2024, when enterprises building on foundation models needed orchestration, memory management, tool access, evaluation, and observability infrastructure that the model providers had not yet built. The middleware companies that stepped into that gap — LangChain, LlamaIndex, Arize AI, Galileo, Pinecone, Composio — attracted meaningful venture capital precisely because the gap was real. What has changed, with gathering speed since mid-2024, is that the model providers have decided the gap represents strategic surface area they intend to own. The question is no longer whether enterprises need harness-layer infrastructure. It is who controls it.

The answer, at the current trajectory, is the model providers themselves. The mechanism differs by company — OpenAI acquires, Google open-sources at scale, Microsoft merges its own frameworks, AWS wraps infrastructure into managed runtimes, Anthropic establishes protocol standards, Meta buys capability at speed — but the direction is uniform. Every middleware function that existed as a standalone commercial offering in early 2024 now has a provider-native equivalent, and in most cases that equivalent is free, deeply integrated, and carries the distribution advantage of the underlying model relationship.

Six Providers, Six Vertical Integration Strategies

The compression is not a monolithic event. Each of the six major providers is executing a distinct integration thesis, and understanding those theses is essential to assessing which middleware positions remain defensible and which have already been absorbed.

The Acquisition Pattern: Three Capability Categories Being Targeted

Across the six providers, the M&A activity reveals three middleware categories that have attracted the most aggressive acquisition interest. These are not random targets. They represent the functions that, if left as independent commercial offerings, would give enterprises the infrastructure to build genuinely provider-agnostic architectures — and in doing so, reduce provider leverage over the enterprise relationship.

1. AI Coding Assistants and Developer Tooling

The Windsurf saga is the defining case study of competitive intensity at the developer tooling layer. OpenAI attempted a $3 billion acquisition that collapsed in July 2025 over Microsoft IP concerns. Google moved immediately — $2.4 billion licensing plus acqui-hire of Windsurf’s CEO and 40 engineers. Cognition AI acquired the remainder for $250 million. The total capital deployed in this single competitive sequence exceeded $2.65 billion. Developer tooling is where model preference is formed at the individual engineer level, and provider competition for that surface area has been correspondingly intense.

2. Evaluation, Security, and Observability

OpenAI’s acquisitions of Context.ai, Neptune, Statsig, and Promptfoo are the clearest expression of this pattern. Each represents a function — evaluation talent, training observability, product analytics, security testing — that enterprises were previously sourcing from independent vendors. AWS added 13 built-in evaluators to Bedrock AgentCore in December 2025. Azure Foundry IQ provides built-in evaluation within Microsoft’s managed runtime. The eval and observability middleware companies that remain independent — Braintrust ($800M valuation, February 2026), Arize AI ($131M raised), Galileo ($68M raised) — are building moats on vendor neutrality and proprietary evaluation IP that providers cannot simply replicate through acquisition of a different target.

3. General-Purpose Agents and Memory

Meta’s acquisition of Manus — the general-purpose agent platform with $100M+ ARR — for more than $2 billion, with Manus continuing to operate independently post-close, represents one of the largest bets on a standalone agent platform in this acquisition cycle. Scale AI ($14.3 billion for 49%, June 2025) provided Meta with data labeling and evaluation infrastructure at a scale no independent vendor could replicate. On the memory side, both Zep AI and Composio reportedly received acquisition offers in April 2025 — within six months of their Series A closes — signaling that strategic buyers recognize memory and integration layers as critical infrastructure worth taking off the board before independent platforms mature.

What Orchestration Frameworks Tell Us About Provider Intent

The most revealing signal in the entire acquisition pattern is what the providers have not acquired: orchestration frameworks. LangChain has raised $260 million at a $1.25 billion valuation. CrewAI has raised $18 million. LlamaIndex has raised $27.5 million. None has been acquired by a major provider. This is not an oversight. It is a deliberate strategic choice by every major provider to build orchestration internally rather than acquire it externally — signaling that orchestration is viewed as core platform capability that providers intend to own and control, not a function they are willing to purchase from an independent vendor and allow to remain neutral.

Microsoft’s merger of AutoGen and Semantic Kernel into a unified Microsoft Agent Framework in October 2025 is the definitive expression of this logic. Two independently developed orchestration approaches were collapsed into a single Azure-centric platform. AWS built Bedrock AgentCore’s Runtime from scratch. OpenAI’s Agents SDK and AgentKit are first-party offerings. Google’s ADK is an open-source project controlled by Google. The orchestration frameworks that remain independent are not independent because providers overlooked them — they are independent because providers decided building was strategically preferable to buying.

The Harness Layer Segmented: What Survives and What Does Not

Not all middleware is equally exposed. The compression affects different harness-layer functions at different rates and with different levels of residual defensibility. Enterprise architects evaluating build-versus-buy decisions and vendor relationships need a clear segmentation of which functions are being commoditized by provider-native offerings and which retain durable independent value.

The Enterprise Strategic Implication

The structural reality that enterprise architects must internalize is that the 79% of OpenAI customers who also pay Anthropic — a figure cited repeatedly in market analyses — are not accidents. They represent a deliberate multi-provider strategy by organizations that have already concluded that single-provider dependence is a governance risk. The Great Compression makes that multi-provider instinct correct but harder to execute: as each provider absorbs more of the harness layer into its platform, the infrastructure required to maintain genuine provider portability becomes more, not less, demanding.

The NIST AI Risk Management Framework’s emphasis on continuous operational risk management — across the full AI lifecycle, not at point-in-time configuration — applies directly to vendor architecture decisions. An enterprise that builds its agentic infrastructure on a single provider’s orchestration framework, evaluation tooling, memory system, and observability platform has not simplified its architecture. It has concentrated its risk. The Great Compression is creating exactly the conditions under which that concentration becomes strategically dangerous: when a provider controls the evaluation layer, it also controls what counts as success for agents running on its infrastructure.

What the Compression Means for Enterprise Decision-Makers

The strategic implications of the Great Compression translate into concrete architecture decisions that enterprise AI leaders — CDAOs, CDOs, CAIOs, CTOs — are making now, with consequences that will compound over the next three to five years as provider platforms mature and middleware vendor consolidation continues.

The first implication is that multi-provider architecture must be an explicit design requirement, not a default assumption. In 2023 and early 2024, multi-provider portability was relatively easy to achieve because the middleware layer was genuinely provider-neutral. As those functions are absorbed into provider platforms, maintaining genuine portability requires deliberate architectural choices — OpenTelemetry instrumentation rather than provider-native logging, vendor-neutral evaluation frameworks rather than platform-built eval, open protocol connectivity (MCP) rather than proprietary tool integrations. Portability is not free. It must be engineered.

The second implication is that open standards favor enterprises over providers in the long run, even when providers sponsor those standards. MCP’s Linux Foundation governance means the 10,000+ servers built by the community cannot be unilaterally deprecated or monetized by Anthropic. A2A’s open-source governance means Google cannot wall off inter-agent communication behind Vertex AI subscriptions. Enterprises should aggressively favor standards-governed infrastructure over proprietary middleware — not because the standards are technically superior in every case, but because governance structure determines who controls the upgrade path, the pricing, and the deprecation schedule.