Count the agents in your GTM stack. Now cut that number by roughly six-sevenths. That's closer to how many are actually agents.

Menlo Ventures put a number on it in its 2025 enterprise survey: just 16% of enterprise AI deployments qualify as true agents. The other 84% are fixed-sequence workflows wearing a new label. Gartner's version of the same finding is blunter — of the thousands of vendors marketing "agentic AI," it estimates only about 130 are real. The rest is agent washing: rebranded assistants, RPA, and scripted bots.

This isn't a semantics argument. The gap between an agent and a workflow is the gap between something that gets better every quarter and something that decays the moment your ICP shifts.

The four-part bar

Menlo's working definition is the cleanest test I've found. A true agent plans, acts, observes feedback, and adapts. Four verbs. Most GTM tools clear two.

Plan and act are easy now. Any tool with a decent model and an API key can decide on a sequence and execute it — draft the email, enrich the record, summarize the call. That's where the category has landed, and it's genuinely useful.

Observe and adapt are where it falls apart. Observing means the system knows what happened after it acted: did the meeting book, did the deal advance a stage, did the rep silently rewrite every line before sending. Adapting means the next run is different because of what it learned.

Almost nothing in GTM does the back half. Your sequencing tool doesn't know which of its emails preceded a closed-won. Your research agent doesn't know that the ICP definition changed in March. They act, and then the loop just... ends.

Why the back half never gets built

Not because vendors are lazy. Because observe-and-adapt isn't a feature you can ship inside a single agent.

To observe, an agent needs to see outcomes that live somewhere else entirely — in the CRM, in the call recording, in the rep's edit. To adapt, it needs somewhere durable to put what it learned, and that place has to be shared, or the learning dies with the tool. A point solution can only ever learn about itself, and only for as long as you keep paying for it.

So the honest architecture of most "agents" is: a prompt, a few integrations, and a fresh start every single time. That is a workflow with better copy.

It also explains the numbers that keep landing on the same desk. Gravitee's survey of 900+ executives found 81% of teams have agents past the planning stage, but only 14% ship them with full security and IT approval — and 88% of organizations report confirmed or suspected agent-related incidents. Adoption is outrunning control, and the gap isn't model capability. It's that nobody can say what an agent did, on whose authority, or whether it worked. A thing that cannot show it improved is very hard to defend at renewal. (More of the underlying data is collected on our agentic AI statistics page.)

The distinction is now a buying problem

For a while this was philosophy. It stopped being philosophy when the count went up.

KPMG's Q2 2026 Pulse found 53% of organizations deploying AI agents, with multi-agent orchestration doubling from 9% to 18% in a single quarter. Teams aren't running one tool anymore. They're running six, from six vendors, each with its own memory, its own permissions, and its own idea of who your best-fit customer is.

Six workflows that each start from zero don't add up to an agentic GTM motion. They add up to six versions of your positioning, six audit gaps, and one very confused RevOps lead trying to explain at QBR why the AI line item should survive.

The buying question is no longer "is this agent any good." It's "what does this agent learn from, and where does that learning live when I switch vendors."

Four questions for your next demo

Take the Menlo bar into the room. It works better than any feature grid.

• Plans: does it decide the steps, or did someone hard-code them? Ask what happens on an input the vendor didn't anticipate.

• Acts: what can it actually write to, and under whose permissions? "Read-only with a copy button" is a draft tool, not an agent.

• Observes: name the outcome it sees. Not opens and clicks — stage movement, closed-won, the rep's edit. If nobody can name the signal, there isn't one.

• Adapts: what is different about run 100 versus run 1, and where is that difference stored? If the answer is "the prompt gets updated by our team," you're buying a consultancy.

Most tools fail on the last two. That's fine — plenty of two-verb tools are worth buying. Just don't pay agent prices for workflow economics, and don't expect a stack of them to compound.

The layer that makes the back half possible

This is the bet we're making at wysdym. Not another agent — the operating layer underneath them.

Bring whatever agents you want: Claude, OpenAI, LangGraph, something your team built. wysdym is the layer they run on: Cortex for shared intelligence and memory, Skills for what agents do, Governance for the rules on every action, Observe for which agents are actually working, and Operator, the included reference agent — all connected through the Gateway, one MCP door to your stack.

Which is really just the four verbs, made structural. Read from a shared graph. Run typed skills. Write through approval queues. Learn from deal outcomes. Put the observe-and-adapt half where every agent can reach it, and a two-verb tool becomes a four-verb one — without the vendor rebuilding anything.

More agents won't grow revenue. Compounding intelligence will.

We're building this with GTM teams already running more agents than they can govern. If that's you, come talk to us.