The Context Layer Is Not the Decision

Data about who a company is (industry, size, who works there) is turning into a utility: one source, many taps, and cheaper every year. This week the vendor with the biggest graph of it (100 million companies) piped that graph into the inbox, chat and CRM tools people already use, bought a company that strings agents together, and shipped agents of its own. That is good news, and it is not the decision. Any automated step has three layers: the context (rented, converging), the signal (what a buyer just did on your site, which only you have), and the decision (your rule, your approval, your record). Own the third. Even the vendor selling the context built bounds, monitoring and cost estimates before it shipped agents, and says the approval step is next.

What happened this week

The week came in three releases from one GTM data vendor (GTM, the sales and marketing side of the company).

On September 29, ZoomInfo launched a connector for Superhuman Go. Its GTM Context Graph ("100 million companies, 500 million contacts, and billions of buying signals," in the release's words) now reads into that assistant through the same interface that already feeds Salesforce Agentforce, HubSpot Breeze, Microsoft Copilot Studio, Gong, LeanData, Claude, ChatGPT and Google Workspace. The vendor's own write-up calls the integration "native and one-way," over API and MCP.

On September 30, ZoomInfo acquired DoubleO.ai, a company that strings agents together. The release says the deal "brings architecture, enabling agents to run complete GTM motions," and gives no terms.

On October 1, it introduced Agent Teams: agents that run GTM plays "from signal to action," available inside its existing GTM Studio "with no new SKU or contract required."

Context became a utility

Read the connector list again. The same graph now flows into the email client, the chat assistant, the CRM copilot, the call recorder, the tool that assigns leads. That is what a utility looks like: one source, many taps. You do not care which pipe the water came through, and soon you will not care which tool the company record came through either. It will be there, current, in whatever you have open.

That is good news for anyone who buys context. Industry, size, stack, the people in the building, who just changed jobs: this used to be a moat and a line item. It is turning into a pipe, and prices for pipes go one way.

It is also the point where a lot of teams draw the wrong conclusion. If the context is everywhere, and the vendor now ships agents on top of it, surely the work is done? No. The context is one of three layers, and it is the one you were always going to rent.

The three layers

Take any step in your stack that runs without a person touching it. Three layers are in play.

Context is who a company is: industry, size, stack, the people in it. You buy it or rent it, and as of this week it arrives through more taps than ever. It is converging. Soon every agent will read the same record.

Signal is what a buyer just did on your site or in your product. A signal: something a buyer just did, like coming back to your pricing page for the second time this week. Nobody sells you that. It happens on pages you own, from accounts you already know or can work out. Two companies with identical context can carry opposite signals on the same morning, and the signal is the part that says "now."

Decision is what happens next and who said it could. A rule you set once: when this happens, do that. Bryn, Civic's agent, calls that a Play. Add the approval that put the rule into force, and the record of each run, written so the person who signs off can read it afterward.

Figure 1. Three layers in any agent-run step: the context converges, the signal is only yours, the decision is the one to own.

The first layer is getting cheap. The second is yours and nobody else's. The third is the one to own, because when an agent acts under your company's name, the question in the room is never "where did the data come from." It is "who decided this, and can I read the run."

We watched a different corner of this in Tools Converge. Action Doesn't., when a data tool added an outreach tool. The short version of that piece: tools keep eating their neighbors, and the action still needs an owner. This one is about context becoming a pipe to everywhere, and about which layer a small team should spend its attention on.

The tell

Here is the part worth reading twice. The vendor that sells the context did not ship agents on the context alone. Read its own account of Agent Teams, published October 1 under the CEO's byline. "Your play and guardrails set the bounds." Each play "defines its triggers, agents, tools, and instructions." Credit usage "is estimated before you activate a play." There is monitoring, so you can "identify and correct mistakes before they affect your business," and role-based access over who may build and run. And in the FAQ, plainly: "A Pause for Review step for human-in-the-loop approval is on the roadmap for a future release."

That is a decision layer under construction, by the company with the largest context layer in the category. Bounds, cost before action, a view of each run, roles, and approval named as the next thing to build. The vendor could have shipped the graph and left the rest to the apps it flows into. It built the layer anyway, because agents that act need one, and it said so. Forkast's read of the launch (October 3) put the reason in one line: "operational risk lives in what agents do, not what they say."

The decision layer is not optional, even for the people who sell the context. That is the tell.

Who decides

For any agent that acts under your company's name, three questions settle whether you own that step or merely host it.

Who decided it could act: a person, a rule a person wrote, or nobody yet?
What bounds did it run inside: one channel, one list, or open?
Can the person who signs off read the run afterward: yes, a log; partly; or no?

Try them on the ordinary steps below. The answers are yours; the line under each step reads them back.

Who decides?

Three ordinary agent-run steps. For each, answer the same questions. The line under each step reads your answers back in plain words.

(illustrative steps; the stack is yours)

Three illustrative agent-run steps and the three questions to answer for each (static view).
Step	Who decided it could act?	What bounds did it run inside?	Can the sign-off read the run?
Send a first email to a new signup	a person / a rule a person wrote / nobody yet	one channel / one list / open	yes, a log / partly / no
Post a pricing-page revisit to the deal channel	a person / a rule a person wrote / nobody yet	one channel / one list / open	yes, a log / partly / no
Add a visitor's company to the outbound list	a person / a rule a person wrote / nobody yet	one channel / one list / open	yes, a log / partly / no

Figure 2. Ordinary agent-run steps and the same questions for each; the line beneath each step restates your answers in plain words (illustrative).

There is no score, because the sentence is the result. "Nobody has decided this step could act yet, so it runs under your name, on an open bound, and nobody can read the run" is a complete finding. We wrote about why the approval is part of the rule rather than a brake on it in Approve Mode Is Not a Speed Bump, and about what the record looks like when the answer is "do not act" in Show Me a Blocked Action.

Where Bryn sits

Bryn watches your site and product for buyers, works out who they are, decides if they're worth acting on, and runs the follow-up you approved, within minutes. Every run is logged, so the people who sign off can read it. Of the three layers, Bryn rents the first: identity and company context come from vendors, and after this week that is a cheaper, better-plumbed layer to rent. Bryn holds the second and third on your behalf: the first-party signal from your own pages, and the decision, which is a Play you approved, a boundary it cannot cross, and a record line for each run. Your own agents can call it too, which is why Bryn is the agent your agents call. Bryn is not another dashboard to watch. It is the governed execution layer that runs Plays through your stack. In plain terms: you write the rule once, Bryn carries it out inside the tools you already use, and nothing runs without a record of who approved it. The follow-up goes out while the buyer is still on your pricing page, and nothing goes out that you'd have to apologize for.

Do this Monday

Take any one automated or agent-run step in your stack. Under it, write three lines: where its context came from, what first-party signal triggered it, and who decided it could run (a person, a rule a person wrote, or nobody). If the third line is "nobody," that step runs under your name with no owner. Fix that one first.

The context layer will keep getting cheaper without your help. The signal is already yours. The decision is the layer to own, and it is three lines of writing away.

Stop watching signals. Start running them.