The Shared Agent Brief: A Written Contract for GTM Teams That Connected AI First

At S2O, three different admins gave the same agents conflicting instructions. Ayush Poddar, who runs the operation, summarized the failure in one sentence: “We had connected the workflow before we had defined the decision rights.” The agents were live, the integrations worked, and nobody had written down who was allowed to decide what.

That ordering error is now common enough to have survey numbers attached. Validity’s State of CRM Data Report 2026, a vendor-run survey of 500 B2B and B2C marketers across five countries at organizations with 100 or more employees, found that 45% already use agentic AI that acts without human review. Two-thirds increased the decisions they delegate to agents in the past year. The data those agents act on is in worse shape: roughly 91% of respondents say data readiness is critical for AI, yet only 21% call their CRM data very well prepared, and only 41% have dedicated data governance owners. Validity sells data-quality tooling, so it is measuring the pain it monetizes. Even discounted for that, the shape of the finding holds: delegation to agents is growing faster than the governance underneath it.

We have written about pieces of this before. AI should not own account selection. A discount guardrail is a decision boundary, not a feature. CRM write access needs action-level scoping. Those were diagnoses and single controls. Poddar’s Shared Agent Brief, published in August 2026, is the first full artifact I have seen for GTM: a written operating agreement a RevOps lead could draft this week. It comes with two honest caveats. The framework is the author’s own practice at one company, with no outcome metrics published yet, and the article carries a sponsored partner mention. Read it as an artifact to adopt and test, not a proven result. The artifact is still worth walking through in detail, because nothing else I have seen in the GTM space is this concrete.

The Five Components

Poddar’s shorthand compresses the whole agreement into one line: shared agent = shared objective + role inputs + decision rights + AI boundary + decision record.

Shared objective. What the agent is for, written once, agreed by every team that touches it. When three admins each optimize the agent for their own function, the agent receives three objectives and silently blends them.

Role inputs. Which humans feed the agent, and in what capacity. For example, sales might feed pipeline context, marketing campaign context, ops process rules. Naming the inputs by role prevents the situation where the agent treats whoever wrote last as the most authoritative voice.

Decision rights. Who is allowed to decide what, in writing. This is the component the S2O incident was missing, and it is the one most teams skip because it forces an uncomfortable internal conversation before any tooling gets configured.

AI boundary. What the agent may do on its own and where it must stop. The boundary is a property of the agreement, written down where every stakeholder can read it.

Decision record. A log of what was decided, by whom, and on what evidence. Without the record, every later dispute about agent behavior degenerates into memory against memory.

None of these components is a product feature. All five are prose that humans agree on before the workflow goes live. Engineering teams already run an equivalent discipline: code review, CODEOWNERS files, merge permissions. The survey numbers suggest GTM wired the automation first; S2O shows what discovering the missing written layer looks like.

The Input Taxonomy: Four Words That Carry Different Authority

The sharpest idea in the brief is a taxonomy of what humans send to a shared agent. Every input is one of four things: EVIDENCE, RECOMMENDATION, DECISION, or APPROVAL. As Poddar puts it, “They don’t carry the same authority.”

The distinction sounds pedantic until you watch a shared agent operate without it. A sales rep pastes a customer quote into the workflow. Is that evidence for the agent to weigh, or a decision the agent must execute? A marketing manager writes that the team should probably pause a sequence. Recommendation or decision? Absent a taxonomy, the agent infers authority from surface signals: seniority of the sender, how recently the message arrived, how imperative the phrasing sounds. Each of those inferences is wrong in a different way.

The taxonomy also settles the conflict case. When two labeled inputs disagree, the brief routes the conflict to a named human conflict owner. The AI never resolves disagreements between humans. That single rule removes the most dangerous failure mode of shared agents: the model quietly adjudicating an organizational dispute that no human knows is happening.

Stop Conditions: Teaching the Agent to Halt

The brief defines explicit conditions under which the agent must stop rather than proceed. Three of them, straight from the artifact: evidence is missing or stale, ownership is unresolved, or no written rule names the decision owner.

The template rule that enforces the last one deserves quoting verbatim: “Do not invent missing policies or decision owners. Mark them UNKNOWN.”

That rule targets the specific way language models fail in operational settings. A model asked to act under an undefined policy will generate a plausible policy and act under it. The generated policy will sound reasonable, will not be flagged as generated, and will differ from what the organization would have chosen. Marking the void UNKNOWN converts a silent fabrication into a visible escalation. It is the GTM equivalent of a null check.

Matched Accounts: Regression Testing for an Operating Agreement

An agreement nobody tests decays into a document nobody reads. The brief’s answer is a matched-account discipline borrowed, in spirit, from software regression testing.

Account B is the clean case: compatible inputs, no conflicts on record. The agent prepares work up to a review draft and sends nothing externally. Account A is the trap: it carries a known conflict that is deliberately absent from the decision record. The correct behavior is to stop, because the stop conditions fire on unresolved ownership.

The pass condition on Account A is the interesting one. The test does not check whether the agent produced good output. It checks whether the agent refused to produce output when the agreement said it must. After any rule change, the brief prescribes five regression checks before the change is trusted. A GTM team that A/B tests subject lines with more rigor than the authority structure governing its agents has the priorities inverted. This flips that.

Why Now

The Validity numbers, vendor-run as they are, describe the exposure. Nearly half of surveyed marketers run agentic AI without human review. Only about a fifth call their CRM data very well prepared, and if those two answers overlap the way the totals suggest, many unreviewed agents are acting on data their operators would not call ready. And 62% of respondents say poor CRM data probably or definitely cost them revenue already. The survey measures self-reported pain, not audited losses, but the direction is consistent across every question: autonomy is scaling, data readiness and governance ownership are not.

The Shared Agent Brief will not fix the data. It fixes the layer above the data: who decides, on what evidence, with what authority, and where the agent must stop when any of that is undefined.

Draft Yours This Week

Do this now: pick the one shared agent workflow in your GTM stack that the most teams touch. Write the five components for it on a single page. Shared objective in one sentence. Role inputs by name. Decision rights by decision type. The AI boundary as a list of allowed actions. A decision record location everyone can read. Add the UNKNOWN rule verbatim. Then build one Account A: a test account with a known conflict deliberately left out of the record, and confirm the agent stops. If it proceeds, you have learned something about your workflow that you would otherwise have learned from a customer.

The document will feel bureaucratic for the first week. So did code review.

This analysis synthesizes The Shared Agent Brief: A Practical Operating Agreement for GTM Teams (StartupGTM, Ayush Poddar, August 2026) and Marketers know AI is using bad data to make decisions (MarTech, Constantine von Hoffman, August 2026).

Victorino Group helps organizations define decision rights, boundaries, and test discipline before agents go live in revenue workflows. Let us talk.