The Great Compression has a new data point.
Anthropic shipped dynamic workflows in Claude Code, and on June 2, 2026 published its account of how they work. The framing was a productivity feature. The event was structural.
The Great Compression series concluded with a complete picture of how model providers are systematically absorbing the agent harness layer — through acquisitions, protocol standards, and managed runtimes. Dynamic workflows adds something categorically different. The harness no longer needs to be acquired or pre-built. Anthropic’s own framing is unambiguous: the title is “A harness for every task,” and the capability is that Claude can now write its own harness on the fly. The orchestration layer stops being infrastructure you bring to the model. It becomes an output the model produces.
Coordination-Grade, Not the Governance Plane
That word — harness — needs disambiguation, because this series has spent considerable effort separating two planes. The Execution Harness is coordination-grade: routing, sequencing, parallelism, the machinery that gets work done. The Agent Harness is the alignment-grade governance plane: the privilege boundaries, the policy enforcement, the independence that lets a system be defended to an auditor. What Claude generates at runtime is the first plane. It is the coordination-grade harness, produced on demand. It is not the governance plane, and it cannot be — governance independence is a structural property, not an inference-time output. Anthropic’s own write-up makes the point for us: among the patterns it describes is quarantine — barring agents that read untrusted content from taking high-privilege actions. That is the governance plane appearing as a pattern a user must think to prompt for, not a property the generated harness carries by default.
The Build Thesis in Product Form
The March 2026 analysis noted that no major provider had acquired LangChain, CrewAI, or LlamaIndex — not as an oversight, but because building was strategically preferable to buying. Dynamic workflows is that build thesis expressed in product form. The patterns that LangGraph and CrewAI built their product surfaces around — classify-and-act, fan-out-and-synthesize, adversarial verification, tournament selection — are named patterns in Anthropic’s own write-up, now reachable through a well-formed prompt. The workflow-construction advantage that justified an orchestration framework is beginning to disappear.
The conventional defense of LangChain’s position has rested on LangSmith — its evaluation and observability product. The reasoning: even if the orchestration layer commoditizes, the observability layer retains independent value. LangSmith does measure across frameworks — OpenAI, Anthropic, custom agents, OpenTelemetry pipelines — so the independence reads as real. But provider-neutral is not orchestration-neutral. The framework-free path gives you generic spans, the same telemetry every OpenTelemetry backend collects. The instrumentation that makes LangSmith more than a generic span logger — automatic run trees, node-level state, native reads of LangGraph execution — originates in LangChain’s own callback architecture. Strip that and the premium goes with it. The differentiated value was built on top of the orchestration framework, not beneath it. When the foundation becomes a platform primitive, the value threaded through it reprices alongside it.
Observability Loses Its Object
There is a deeper pressure than coupling. Observability of orchestration assumed a stable object to observe — a workflow you instrument once, watch fail in production, correct, and redeploy. Dynamic workflows dissolve that object. When the orchestration is generated per task and discarded when the task is done, there is no persistent artifact to monitor across runs and no durable target to improve. You can still trace a single execution. You cannot optimize a workflow that does not outlive the prompt that produced it. The question stops being whether the observability layer can see the workflow. It becomes what stable thing is left to see.
The broader segmentation from the series still holds: vendor-neutral evaluation with proprietary models, cross-provider observability built on OpenTelemetry, and governance infrastructure that requires independence from the model provider by definition — these retain defensible value precisely because their business model is not structurally conflicted with provider-neutral measurement. What dynamic workflows compresses further is the middle: the workflow construction, routing, and orchestration logic that required a framework in 2023 and 2024, and that now requires a well-formed prompt.
The timing matters as much as the capability. Dynamic workflows arrived less than six months after Claude Code established the execution runtime, and less than a year after Agent Skills defined the four-layer harness architecture. The sequencing is deliberate. Each release is a land claim. Individually they look like product features. In sequence they describe the systematic completion of a native harness that leaves no gap for third-party orchestration frameworks to fill.
The series described three mechanisms of absorption: providers acquire the tooling, providers set the protocol, providers operate the runtime. Dynamic workflows introduces a fourth — providers produce the coordination-grade harness: generated at inference time, custom-fit, discarded when the task is done.
Once orchestration becomes a model output rather than a software artifact, the standalone orchestration framework loses its reason to exist.
