AI agent orchestration: who owns the plan when every tool has its own agent?

I’ll be honest, I didn’t set out to write about orchestration. I set out to figure out whether our own product had a category.

If you’re a marketing leader in 2026 you already know the shape of the problem. HubSpot has Breeze. Salesforce has Agentforce. Adobe rebranded Experience Cloud as CX Enterprise and put Coworkers at the centre of it. Microsoft and Anthropic launched Copilot Cowork together. And on 1 May, Microsoft Agent 365 goes live as the control plane for all of them. Your marketing stack is about to contain ten or more agents from five or more vendors, each good at its own slice of work, each shipping on its own release cadence, none of them talking to the others.

AI agent orchestration is the coordination layer that sits above all of that. It holds a persistent plan. It delegates to task-shaped agents. It follows up when things slip. Without it, you have a room full of smart helpers and no one in charge.

That matters more than the agents themselves. A recent Salesforce report found that large enterprises now run an average of 957 applications, but only 27% are actually integrated. Roughly half of AI agents already live in silos, disconnected from the broader marketing stack (ConvertMate analysis of Salesforce 2026 Connectivity Report ). The agents are the easy part. The plan ownership is the hard part.

I’ve been building Fifty Five and Five for eleven years, and for the last two of them we’ve been quietly building an orchestrator. I’ll say upfront: this post is about the category, not a pitch. If the category didn’t exist, we wouldn’t have a product in it.

What is Copilot Cowork? A plain-English explainer

Copilot Cowork is the most interesting launch of the last 60 days because it proves the pattern. So let me define it properly before anything else.

Cowork matters because it is the first time a hyperscaler has credibly shipped long-running, multi-step agentic work at enterprise scale. It is a genuinely good product. Microsoft says 80% of Fortune 500 customers already deploy its AI agents, and internally they can see more than 500,000 of them running across customer tenants (Fortune, 9 March 2026 ). The adoption is real.

But Cowork answers requests. It does not own plans. You tell it “prep me for this meeting” and it does. The next time you open it, it does not remember what your quarterly goals are or whether the pipeline number you showed it three weeks ago has moved. That is not a bug, that is the design. It is a task engine.

The rest of this piece is about what sits above it.

Mapping the enterprise AI agents landscape in 2026

I’ve been tracking these launches for two years and I have lost count. Let me try to put the main players on one page.

Each of these is good at its own work. None of them owns a plan that spans more than one of these columns. None of them knows that last week’s campaign hit its CTR target and this week’s has not, unless that data happens to sit inside the same vendor’s walled garden.

Here’s the thing nobody talks about. You’re not choosing between these. You’re getting all of them. HubSpot Breeze ships with the CRM you already pay for (HubSpot Spring 2026 Spotlight ). Agentforce Marketing is bundled into your Salesforce licence. Adobe Coworkers come with Experience Cloud. And on 1 May, every Microsoft 365 E7 seat in your company becomes a Cowork user whether your CMO signed off on it or not.

Why an AI marketing platform is not the same as an AI agent stack

I get asked almost weekly which AI marketing platform a team should buy. It is the wrong question.

A platform is one vendor’s model of how work should flow, with AI features layered on top. You buy it and you commit to their shape: their data model, their integrations, their idea of what a campaign looks like. That is fine when you want one coherent experience for one team.

An agent stack is the opposite shape. It is many agents from many vendors, each doing a different piece of work, loosely federated. That is not better or worse, it is a different problem with a different answer.

The numbers back this up. ConvertMate’s analysis of B2B marketing teams this year found that those running more than eight disconnected AI tools lose about 34% of their AI spend to redundancy alone (ConvertMate, 2026 ). Half of the AI agents enterprises already run operate in silos, disconnected from the rest of the marketing stack. That is what a stack without an orchestrator looks like.

You cannot fix that by buying a bigger platform. A platform cannot tell an agent inside another vendor’s cloud to wait, slow down, or change tack. Only an orchestration layer can do that.

So when someone asks me which AI marketing platform to buy, my honest answer is: you are probably going to have five of them whether you like it or not, because they are bundled with the tools you already run. The real question is what sits above them.

The missing layer: an AI chief of staff for your plan

Here is the analogy that finally made this click for me.

A good human chief of staff does not do the work. They own the plan. They know what is due next week, who is supposed to ship what, and where the risk lies. They chase people when things slip. They surface what the CMO actually needs to see. They make other people’s work possible.

An AI chief of staff is the same shape. It does not replace Cowork, Breeze, or Agentforce. It sits above them.

That requires four things, and this is where most of the current crop of “orchestration” tools fall down.

1. A persistent plan. Not a request-and-response assistant. Something that remembers the goal you set six weeks ago, the constraints you agreed to, and the status of every sub-objective since. In our own product we split this into three tiers: master plan, tactical plan, and weekly plan, each deriving from the one above.

2. Proactive triggers. Waking up when a deadline slips, a metric drifts, or a milestone completes, without waiting for a prompt. Request-response agents never do this. The whole point is that the plan owns the agent, not the other way around.

3. Delegation across vendors. Task agents live in different clouds. An orchestrator needs to route work to whichever one is right for the job. The trick is to stop waiting for every agent to expose a neat API. If an agent has a user interface, it is addressable. Anthropic donated MCP to the Linux Foundation in December 2025 and the Agent2Agent (A2A) protocol followed in January 2026, which covers the agents that do expose themselves properly. For the rest, the answer is browser automation. Services like Anchor Browser let an orchestrator drive any web UI under a real user session, with 2FA and SSO handled on the side. So an agent you cannot call via API you can still call via its own front end. Either way, the orchestrator treats it as a tool.

4. A surface people actually use. It has to live somewhere your team already checks. For most enterprise teams that is Microsoft Teams. A beautifully engineered orchestrator that no one opens is worse than no orchestrator at all.

This is what we have been building with Compass for the last two years. I will not pretend it is a marketing specialist, because it is not. Under the hood it is a general-purpose strategic planning and execution agent that lives in Teams, holds plans, and delegates across tools via MCP and browser automation. That is the important bit. To Compass, an AI agent is just another tool with an interface. If the agent exposes MCP or A2A, we call it directly. If it only exposes a web UI, we drive that UI under a real user session with 2FA relayed through Teams. Marketing is the first shape of work we have taken it to because that is the shape of work we understand best . The engine is domain-neutral, and new territory. We are adding agents one at a time, each with a tested adapter. That is the only honest way to build this.

Deloitte’s 2026 Technology, Media and Telecom predictions put the gap in numbers: only 28% of executives say their organisation has mature AI agent capabilities, compared with 80% for basic automation (Deloitte, 2026 ). That gap is not about the agents. It is about the layer above them.

What AI agents for marketing should do next week

If you take one thing from this post, take this: do not try to solve orchestration in one go. Start small.

Here is the sequence I would use with a client this month.

Pick one plan you’d like owned end to end. Not your whole marketing function. One plan. A launch. An ABM programme. An SEO and GEO programme. Something with a clear north star, a timeline, and measurable outcomes.

Inventory the task-shaped agents you already have. Cowork comes with E7 from 1 May. Breeze comes with your HubSpot seats. Agentforce Marketing comes with Salesforce. You almost certainly own more AI than you think you do.

Pick an orchestrator. Three options. Build one, which is possible, expensive, and slow. Buy one, and more of these exist every month, including ours. Or wait for the bundled one. Microsoft Agent 365 on 1 May is the obvious candidate and worth watching, but note that Microsoft is marketing it as a governance and observability plane, not as a plan owner. Governance is the floor, not the ceiling.

Apply the plan test. Ask any candidate orchestrator one question: does it remember the plan between conversations? If it only answers prompts, it is a task agent. There is nothing wrong with that, but it is not what you need for this job.

We have seen the pattern work in simpler forms already. An AI-powered blog writing process, for instance, chains together keyword research, synopsis creation, draft writing, editorial review, and publication. Each step is its own LLM-backed component with a specific job. Taken together they are a small example of the orchestration pattern: a plan that sits above the pieces, not inside any one of them. The same pattern scales to commercial agents. You do not need Cowork or Breeze to expose a special API for you. You need them to have a front end, which they do. An orchestrator with browser automation can drive any of them the way a user would, which is where Compass is heading next.

The scaling gap is real and bigger than most leaders realise. A recent analysis by Digital Applied found that 78% of enterprises have AI agent pilots running, but fewer than 15% have scaled any of them to organisation-wide operational use. Only 14% have a single agent genuinely running in production at that level (Digital Applied, March 2026 ). The pattern across every client I’ve seen is identical: the agents work in demos, then hit the wall because nothing owns the plan they are supposed to execute against.

Start with one plan. Layer an orchestrator on top. Plug in three task agents. Ship it. Then expand.

Who owns the plan? A recap for marketing leaders

The question I started with was “who owns the plan when every tool has its own AI agent?” The honest answer is that today, nobody does, by default. The agent layer has matured fast. The layer above it is still being built.

Here is what to take away.

• Horizontal task agents are real and worth using. Copilot Cowork, HubSpot Breeze, Salesforce Agentforce, and Adobe Coworkers will all be part of your stack within twelve months whether you buy them deliberately or inherit them.
• None of them owns the plan. That is by design, not oversight. They were built to execute tasks on request, not to hold multi-week strategic plans.
• AI agent orchestration is the missing layer. Think of it as an AI chief of staff: persistent plan memory, proactive triggers, delegation across vendors, and a surface your team actually uses.
• Do not wait for “the one AI marketing platform to rule them all”. It is not coming. Design an agent stack and layer an orchestrator on top instead.
• Start with one plan, one orchestrator, three task agents. Prove the pattern, then expand.

If you want to talk about where an orchestrator fits in your stack without betting the farm on it, get in touch . I’m genuinely interested in what plans you are trying to own, and what agents you are already paying for that could plug into one.

Frequently asked questions