ANALYSIS The monolithic AI agent is breaking apart. Across open-source releases, enterprise playbooks, and practitioner builds, the industry is converging on a three-layer separation: harness, framework, and protocol, each owning distinct responsibilities. The split is not academic taxonomy; it is driven by the economics of context windows and the operational reality that agents fail at state, permissions, and recovery, not at inference.
Why it matters
OpenAI open-sourced the Codex Harness framework under the Apache-2.0 license on August 20, releasing the codex exec CLI tool, the Codex SDK, and the app-server7. Anthropic's Claude Code docs describe the same architectural concept as an "agentic harness," with the Claude Agent SDK exposing "the same tools, agent loop, and context management that power Claude Code"3. DeepSeek released DeepSeek Harness v0.1 on August 13 under the MIT license2. ◆ Three separate labs shipping harness-layer code within days of each other is not coincidence; it reflects a shared conclusion that the execution wrapper around a model is a distinct, separable product surface.
The big picture
The taxonomy now crystallizing treats each layer as owning a different set of concerns. The harness is "the execution system that wraps a model and turns it into a working agent," managing conversation state, streaming execution, tool use, sandbox enforcement, and approval policies, as OpenAI's Codex-as-a-platform post defined it. A framework, by contrast, is "a library of primitives for composing agents," covering model clients, tool abstractions, graph orchestration, memory interfaces, and observability hooks; examples include LangGraph, the OpenAI Agents SDK, and Microsoft Agent Framework, which reached 1.0 GA in April 2026. MCP, the Model Context Protocol, sits below both as "a wire protocol, not a runtime," using JSON-RPC 2.0 messages to standardize how an LLM application discovers and calls capabilities exposed by servers.
The protocol layer is maturing fast. The 2026-07-28 MCP revision retired the initialize/initialized exchange and the Mcp-Session-Id header, deprecated server-initiated sampling, roots, and logging, introduced Multi Round-Trip Requests as their replacement, and deprecated the legacy HTTP+SSE transport with a 12-month offramp. Since December 2025, the Linux Foundation's Agentic AI Foundation has governed MCP alongside Goose, AGENTS.md, and A2A. The TypeScript and Python MCP SDKs have each passed 1 billion total downloads.
Between the lines
The real pressure behind this decomposition is cost, not architecture aesthetics. Jeff Hollan, VP of PM for Microsoft's Foundry Agent Service, wrote that in many production systems the context-window strategy "was set during prototyping and never revisited, even though it often drives the largest share of operating cost and contributes to disappointing answers"5. Microsoft's framing treats context engineering as the lever that makes "each agent cheaper over time as it learns what works".
OpenAI's own data puts a number on the harness payoff: optimizing Harness design alone boosted GPT-5.6 Sol's score on the ARC-AGI-3 benchmark from 13.3% to 38.3% while cutting token consumption sixfold6. Thrive Holdings and Cisco have deployed the framework for tax filing and cloud management, processing 7,000 forms and increasing efficiency by roughly one-third.
Practitioners are building toward the same separation independently. Jeff Ng, a founding engineer at Unblocked, argued that "building agents has actually gotten pretty easy, but unfortunately they still get things confidently wrong," pointing to checkpoint and state persistence as a core unsolved problem: "Crashing without durability can actually lead to a lot of state loss" including message history, tool calls, and loop position1. One developer built a "retrieval + orchestration control plane" that separates task routing, retrieval routing, context assembly, structural analysis, execution, and verification into distinct stages, with budget-constrained context selection and evidence-safety gates that can prevent an agent from editing when evidence is weak10. Another open-sourced ContextOS, a token-budget-aware context orchestration layer for long-horizon agents that treats retrieval and context selection as separate problems11.
ANALYSIS The pattern across these efforts is consistent: practitioners who operate agents at production scale find that the model call is the cheapest problem to solve, while state management, context budgeting, and recovery logic are where failures and costs concentrate. The layered architecture lets teams swap or upgrade each concern independently.
The debate is also moving beyond loops. In mid-2026, Peter Steinberger of OpenClaw posted on X: "Are we still talking loops or did we shift to graphs yet?" prompting AI engineer Hamel Husain to publish an article titled "Loop Engineering Is Dead. Enter Graph Engineering"4. Codex hooks, documented by Shuai Guo, let developers attach custom logic at lifecycle events such as SessionStart, PreToolUse, PostToolUse, and Stop, effectively inserting graph-like control points inside the harness loop9.
What's next
OpenAI's Harness is now Apache-2.0 licensed and shipping three components that any team can embed. Microsoft's four-part Economics of Agent Optimization series is still publishing, with a fourth post forthcoming. The MCP specification's 12-month deprecation clock on legacy HTTP+SSE transport sets a concrete migration deadline. The MCP maintainers report close to half a billion SDK downloads per month across Tier 1 SDKs, a volume that will pressure every harness and framework to treat the protocol layer as a first-class integration surface rather than an optional adapter.