The hottest thing an enterprise can buy in the third quarter of 2026 is a place to control its AI agents. On September 1, a London startup called Orchestra launched its “Agentic Control Plane” with $4.6 million in fresh funding, landing on top of a category IBM had already recast watsonx Orchestrate around and Deloitte had already written a playbook for. An AI agent control plane is the layer that deploys, monitors, and governs a fleet of agents from one console — who can run, what tools they touch, what gets logged. It is a genuinely useful thing to own. It is also the wrong half of the problem. A control plane can tell you your agents are behaving. It cannot tell you what they actually shipped. That second question — what got done, by whom, and whether it is truly finished — belongs to a different layer, one the 2026 discourse hasn’t named yet. Call it the work plane. Lova is a chat-first AI project management product where AI agents work as teammates: they claim bounded tasks on a shared board, move them through explicit states, and leave an audit trail every teammate — human or agent — can read. That is the work plane. Not a console for governing the machinery. A surface where the work itself lives.
Key takeaways
- The “agent control plane” is the enterprise land-grab of 2026. Orchestra launched one on September 1 on $4.6M; IBM and Deloitte are already in the category.
- A control plane governs how agents behave — orchestration (how they work) and governance (whether they should). Both face down, toward the machinery. Neither faces up, toward the work.
- The fear driving the buy is real: Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027 — on cost, unclear value, and weak controls, not on model capability.
- The evidence says the missing piece is organizational, not infrastructural. Microsoft’s 2026 Work Trend Index finds organizational factors drive about 67% of AI’s real impact versus 32% for individual skill.
- The fix isn’t a smarter model or a deeper control plane. It’s a work plane: a shared board where every task has an owner, an explicit state, and a “done” that has to be earned — readable by humans and agents alike.
What is an AI agent control plane?
An AI agent control plane is a centralized system for operating a fleet of agents: deploying them, watching what they do, and enforcing the rules that keep them in bounds. IBM’s version, which it built into watsonx Orchestrate, pitches itself as “one place to operate, govern, and scale AI agents” from any framework — end-to-end visibility into what agents are doing, built-in controls over how they behave, a catalog to reuse them. The best short definition of the category comes from the way its own proponents split it: a control plane distinguishes orchestration — how an agent performs work — from governance — whether it should be doing that work at all, with which tools, on whose data, recorded for which audit. Both are worth having. Both, notice, describe the agent’s behavior, not the work’s outcome.
This is why the category is booming. In a year when a company might onboard hundreds of agents from a dozen frameworks, running them without a control plane is like running a data center with no admin panel. Orchestra’s September launch is the fresh face of a real need: after growing usage more than 10x, it raised a $3.3 million seed led by Differential Ventures and a $1.3 million pre-seed led by Moonfire to bring “pipelines and governed AI agents into one environment.” The pitch writes itself, and it should. But every product in this category answers the same question — are my agents under control? — and leaves a second one wide open.
Why is every enterprise buying an AI agent control plane in 2026?
Because the alternative is failure that shows up on a budget line. Gartner’s forecast — that over 40% of agentic AI projects will be scrapped by the end of 2027 — has hung over every agent budget this year. The reasons Gartner names are telling: escalating costs, unclear business value, and inadequate risk controls. Model capability isn’t on the list. The same analysts estimate only about 130 of the thousands of vendors waving the word “agentic” are the real thing, which means a lot of the spend is chasing a label. A control plane is the natural reflex: if projects die from weak controls, buy controls.
And the fleets keep growing regardless. McKinsey’s State of AI survey finds that among companies above $1 billion in revenue, 40% now report scaling AI agents somewhere in the business, up from 27% a survey earlier. Microsoft’s 2026 Work Trend Index clocks a 15x jump in active agents inside its 365 suite year over year — 18x at large enterprises. When you have that many agents acting on your behalf, wanting a console to govern them is not paranoia. It’s hygiene. The mistake is thinking the console is the whole job.
Control plane vs. work plane: what’s the difference?
Here is the frame the industry is missing. Borrow the language from infrastructure, where a control plane sits above a data plane and decides how the machinery runs. The agent control plane keeps that shape: it sits above the agents and governs the machinery — the runs, the tools, the permissions, the logs. It faces down, toward execution.
A work plane faces the other way. It sits above the agents too, but it points at the work: the discrete, ownable units a team is actually trying to finish. On a work plane, a task has a name, a single owner, an explicit state it can be in, and a definition of done that has to be satisfied before it moves. The control plane answers “is this agent allowed to act, and did it behave?” The work plane answers “what are we shipping, who has it, and is it actually finished?” You can pass every question on the first plane and have no honest answer to a single question on the second. A fleet of perfectly governed agents can still deliver a pile of half-done work that no one owns and no one can verify.
This distinction isn’t academic. It’s the difference between knowing your agents are safe and knowing your project is real. The whole control-plane conversation of 2026 has collapsed “we can see the agents” into “we can see the work,” and those are not the same sentence. We made the narrower version of this argument when we said agents need a system of record, not a group chat. The control plane is the enterprise-grade version of the same category error: a beautifully instrumented view of behavior, with the work itself still nowhere in particular.
Why can’t a control plane tell you what your agents shipped?
Because it was never pointed at the work. Governance sits, by design, outside the execution loop — the whole argument for a control plane is that it should be independent of the tools agents run on, so it can watch them objectively. That independence is exactly why it can’t see outcomes. It observes actions: a call was made, a tool was used, a policy held. Whether those actions added up to a finished, correct, owned piece of work is a judgment that lives one layer up, in the work itself — and the control plane, sitting outside, has no view of it.
The data keeps pointing at that upper layer. Microsoft’s Work Trend Index finds that organizational factors — culture, manager support, how work is structured — account for about 67% of AI’s real impact, against 32% for individual skill, and that only 26% of workers say their leadership is clearly and consistently aligned on AI. McKinsey’s reading is the same from the other side: the high performers pulling real value from agents are nearly three times as likely to have fundamentally redesigned their workflows, not just bolted agents onto the old ones. Both findings say the payoff hides in how work is organized — the exact thing a control plane, watching the machinery, is built not to touch. Buying more control-plane doesn’t move the two-thirds. It polishes the third.
Where the work actually lives
A work plane is a shared board, and it’s worth being precise about the entity. Lova is a chat-first AI project management product built around a board where AI agents are first-class teammates. You steer in plain language — that’s the chat-first part — but every instruction resolves into a change on the board underneath: a bounded task, claimed by a named owner, moved through an explicit state, with a trail written behind it. The chat is the steering wheel. The board is where the work becomes visible, ownable, and finished.
Read against the control plane, the two layers stop competing and start composing. Let the control plane do its job: decide whether an agent may act, with which tools, on whose data, and keep the log. Then let the work plane do the job the control plane can’t: hold the task, name the owner, gate the “done,” and show a human exactly what got shipped. An agent doesn’t announce in a channel that it finished; it moves a card, and the move is the signal. A second agent doesn’t collide on the same work, because the claim is a lock, not a message. This is the same reason we keep arguing the enterprise runs on a board, not an org chart — and the reason that, without one, nobody actually manages your agents, no matter how many consoles watch them. A control plane makes a fleet safe. A work plane makes a team.
Why does this matter now, in Q3 2026?
Because the budgets are being written this quarter, and the category on the invoice is “control plane.” The land-grab is on — Orchestra in September, IBM and Deloitte ahead of it, a dozen more behind — and the framing is winning before anyone has noticed what it leaves out. If 40% of agentic projects really are headed for the scrap heap by 2027, the survivors won’t be the ones with the tightest console over their agents. They’ll be the ones who also gave the work somewhere to live: a plane where a task has an owner, a state, and a done that’s earned, whether the teammate who claims it is a person or an agent. Govern the fleet. But don’t confuse a well-behaved fleet with a finished project. Those are two planes, and most of 2026 is only buying one.
Frequently asked questions
What is an AI agent control plane?
An AI agent control plane is a centralized layer that deploys, monitors, and governs a fleet of AI agents from one place — setting who can run, which tools and data are permitted, what is inspected in real time, and what is recorded for audit. It typically distinguishes orchestration (how agents perform work) from governance (whether they should be doing it at all). Both describe agent behavior rather than the outcome of the work.
What is the difference between a control plane and a work plane?
A control plane sits above the agents and faces the machinery: runs, tools, permissions, logs. A work plane sits above the agents and faces the work: discrete tasks with an owner, an explicit state, and a definition of done. The control plane answers “are my agents behaving?” The work plane answers “what got done, by whom, and is it finished?” You need both, but only the work plane can tell you what your agents actually shipped.
Do AI agents still need a control plane if they have a shared board?
Yes — the layers compose rather than compete. A control plane decides whether an agent may act and keeps the audit log; a shared board (the work plane) holds the task, names the owner, gates the “done,” and shows humans what shipped. Governing the fleet without a place for the work leaves you safe but blind to outcomes; a board without governance leaves you clear on outcomes but loose on control. Real teams want both.
What is Lova?
Lova is a chat-first AI project management product built around a shared board where AI agents work as first-class teammates. You steer the work in plain language, and every message resolves into a change on the board: a bounded task claimed by a specific owner, a status moved through an explicit state, a trail written. Because every action becomes a visible, recorded transition, humans and agents stay aligned on what was actually delivered — the board is the work plane the control plane can’t be.
Why are 40% of agentic AI projects expected to fail by 2027?
Gartner attributes the projected cancellations to escalating costs, unclear business value, and inadequate risk controls — not to weak models. Each of those is a coordination and outcome problem more than a capability problem: work with no clear owner, value no one can point to, and behavior no one can verify. A control plane addresses the risk-control piece; the cost and value pieces live on the work plane, where results are made visible and “done” has to be earned.