ANALYSIS AI coding agents have crossed a threshold: they no longer just write functions on command but orchestrate entire software lifecycles, from schema design through deployment, while the developer steps away. That shift is already reshaping enterprise procurement, coordination architectures, and the tooling layer that sits between agents and production infrastructure.
Why it matters
Nearly 32 percent of organizations have decided against buying one or more software products because they could be built internally with agentic coding tools, according to McKinsey's "The state of AI in 2026" report6. Technology companies lead at 41 percent, followed by healthcare payers and providers at 39 percent and professional services and energy and materials each at 38 percent. Financial institutions register 36 percent, media and telecom 34 percent, and pharmaceuticals 33 percent. ◆ Those numbers represent a structural demand-side shift: when a third of buyers can plausibly replace purchased software with agent-built alternatives, every SaaS vendor's renewal pipeline carries new risk.
The big picture
The infrastructure enabling this shift is maturing on several fronts simultaneously. OpenAI open-sourced Harness, the core execution framework powering its Codex coding agent, under the Apache-2.0 license on August 2014,13. The release includes three components: the codex exec CLI tool, the Codex SDK, and the app-server, supporting persistent conversations, real-time streaming, and human approval workflows. OpenAI's own data shows that optimizing Harness design alone boosted GPT-5.6 Sol's score on the ARC-AGI-3 benchmark from 13.3 percent to 38.3 percent while cutting token consumption sixfold. Thrive Holdings and Cisco have already deployed it for tax filing and cloud management, processing 7,000 forms and reducing preparation time by roughly one-third.
Meanwhile, the coordination problem is being attacked directly. Researchers at Coral AI Labs introduced AgentRadio, an asynchronous message-passing layer that lets agents communicate between execution steps without interrupting their main work4. On a benchmark of long-horizon questions over production repositories, a team of four agents powered by AgentRadio nearly doubled task accuracy compared with four Claude Code agents working independently, and outmatched single agents running on more advanced models.
The deployment gap is narrowing too. AppWizzy launched Projects MCP, an MCP server that enables AI coding agents to create, deploy, and monitor cloud projects through a secure OAuth connection5. An agent can now take code from a workspace, create a project, select a machine, and monitor the project until it is running. A free open-source utility called Agent Island watches local Claude Code and OpenAI Codex sessions from the menu bar, alerting developers when an agent finishes, errors out, or needs human input8.
Between the lines
Jeff Ng, a founding engineer at Unblocked, framed the real constraint plainly: "Building agents has actually gotten pretty easy, but unfortunately they still get things confidently wrong"3. Six months ago, he noted, building an agent "required a team's effort and basically a quarter" because each supporting system, from checkpoint persistence to state management, "was basically its own company or at least a company function". Several frameworks now target that complexity: Analytics Insight catalogued options including LangGraph for stateful workflows with checkpoints and human-in-the-loop processes, OpenAI Agents SDK, CrewAI, Mastra, and PydanticAI for different developer needs, and Google ADK and Microsoft Agent Framework for cloud and enterprise applications11. ANALYSIS That proliferation compresses the timeline Ng described, but his observation points to where value is migrating: away from the agent loop itself and toward the context layer that prevents confident errors.
The New Stack captured this in its taxonomy of three agent roles inside developer platforms: agents as platform consumers that pull service ownership, dependencies, and standards before writing code; agents as platform operators; and agents as platform builders2. "A lot of companies said they treat their AI agents like employees, and in a development platform that makes the agent just another engineering resource consuming it," the analysis noted. ◆ That framing explains why the McKinsey build-vs-buy numbers skew highest in technology and healthcare, sectors with mature internal platforms that can feed agents the context they need.
NVIDIA's Holoscan team ran an ablation study showing that providing a coding agent with the combined resources of the Holoscan CLI, skills documentation, and examples led to the most efficient and highest-quality development workflow compared with using documentation or CLI alone15,1. ◆ The result reinforces the same pattern: raw model capability matters less than the structured context surrounding it.
One solo developer building a mobile app described the practical consequence: he wrote the full database schema before touching a single screen, reasoning that "a screen is cheap to redo" but "a wrong data model discovered halfway through building screens is a different kind of expensive"7. ◆ Agent-driven development does not eliminate architectural judgment; it concentrates it at the front of the process.
What's next
The share of larger enterprises scaling agents in one or more functions rose from 27 percent to 40 percent, while smaller organizations remained flat at 22 percent. ◆ That divergence suggests the build-vs-buy shift will accelerate unevenly, favoring organizations with the platform maturity to supply agents with reliable context. OpenAI's Harness release under Apache-2.0 lowers the framework barrier; the next bottleneck is the service catalog, dependency graph, and institutional knowledge that agents need before they can replace a purchase order.