A meta harness — software that composes multiple AI agents into one shared system — is being reinvented in parallel by independent teams. Who owns that layer will decide which AI tools plug into which enterprise systems.
A dozen teams, ranging from startups to hyperscalers to open-source labs, are independently rebuilding the same layer of AI infrastructure at the same time. The work goes by different names in different repos, but the pattern is consistent: somebody needs to sit above the agent and orchestrate the agents below it. The question is no longer whether the layer exists. The question is who, if anyone, ends up owning it.
The current window has a name. The AI newsletter Latent Space used the phrase "Meta-Harness Summer" in its June 23–24, 2026 roundup to describe a wave of orchestration layers positioned above individual coding and knowledge-work agents, an umbrella over products like Anthropic's Claude Code, OpenAI's Codex, and the agent SDKs that ship from model labs. The newsletter traces a short lineage: Conductor and Zed's ACP, OpenInspect, Cloudflare's Flue, Vercel's Eve and HarnessAgent, and Heypi, each tackling the same problem from a different angle. Latent Space's read is that roughly 1,000 AI-native shops are rediscovering this pattern in parallel, a framing worth quoting but not quantifying, since the count is editorial, not measured.
To see why the layer is contested, start with the term "harness." In agent engineering, a harness is the runtime that wraps a foundation model and gives it tools, memory, file access, and a loop to act in. Claude Code, Codex, and similar products are themselves harnesses. A meta-harness is one level up: software that composes multiple harnesses into a single system, switching between them, sharing state, and applying governance rules across the whole stack. Databricks' open-source Omnigent release is the most prominent recent example. Matei Zaharia of Databricks announced the project on June 13, 2026, describing it as a "meta-harness for AI agents" that combines Claude Code, Codex, Pi, and agent SDKs with live collaboration and a policy-driven control layer. The project's public GitHub repository ships a dedicated Policies document for that governance surface, and a product site markets it as pluggable, secure, and scalable. Independent tech press Heise and Efficiently Connected covered the release in the same framing.
The reason every shop is reinventing the layer comes down to enterprise plumbing. Most production agent systems do not run a single model through a single harness. They run a coder agent, a research agent, a retrieval agent, and a tool-using agent, and they need shared memory, audit trails, rate limiting, and policy enforcement across all of them. Each of those cross-cutting concerns becomes a wheel that every team has quietly reinvented. Omnigent, according to Databricks' own blog post, is positioned as a pluggable, standardized answer to that wheel-reinvention problem.
The closest prior art is the Model Context Protocol, the open standard that lets AI tools call external services in a uniform way. MCP did not win because Anthropic pushed hardest. It won because the problem it solved, letting any model reach any tool, was general enough that every major lab eventually agreed on the wire format. The meta-harness layer is asking the same question one rung up: who decides the format for composing agents with shared governance? If the answer is a single vendor's repo, lock-in follows. If the answer is an open protocol, the layer becomes infrastructure, the way HTTP did for the web. As Latent Space put it in the roundup that named the moment, it remains "unclear whether or not" any of the current contenders will converge.
A few honest caveats belong in the picture. Omnigent shipped roughly eleven days before the newsletter's framing of "Meta-Harness Summer" went out, which makes any adoption claim premature; Heise's coverage treats it as a release, not a market position. The Latent Space "1,000 AI-native shops" line is rhetorical, not a sourced count, and the newsletter itself flags it as editorial. Some of the named lineage, including Conductor, Zed's ACP, OpenInspect, Flue, Eve, HarnessAgent, and Heypi, has been characterized in the same paragraph rather than independently corroborated, so the pattern is the story, not the roster. And a meta-harness can also be read as MCP rebranded for a different layer of the stack, a real possibility the source itself does not exclude.
What to watch over the next two quarters: whether any of the named projects adopts an open protocol that the others can interoperate with, and whether the major model labs treat the orchestration layer as a feature of their own SDKs or as a neutral substrate. The first sign of consolidation will not be a product launch. It will be a pull request.