06-reference/research

enterprise agent roles vs agent pets default

2026-07-04·research-brief·source: deep-research
agent-rolesagent-deployerenterprise-aicompetitorcompounding-intelligence

The Enterprise Agent Default Is a Vote — and the Big Three Are Shipping Centrally-Owned Agent-Roles, Not Personal Pets

The question

What are Salesforce Agentforce, Microsoft Copilot for M365, and Google Workspace (Gemini/Duet AI) shipping BY DEFAULT in 2026 — the "agent-roles" pattern (shared team agents with defined jobs, owned/administered centrally) or the "agent-pets" pattern (per-employee customized personal copilot)?

Context: Every's 2026-05-15 retraction of the personal "Plus-One" model in favor of shared team-agents is the cleanest published operator evidence that personal-pet agents don't scale in orgs. This brief tests whether the three biggest enterprise-productivity vendors' shipped defaults corroborate that — distinguishing what each ADMINISTERS centrally (roles) vs what each end-user CUSTOMIZES (pets).

What we already know (from the vault)

What the web says

Salesforce Agentforce — clearest roles-default of the three.

Microsoft — a deliberate two-layer split: personal Copilot (thin pet) + centrally-governed custom agents (roles).

Google Workspace — most pet-friendly on authoring, still roles-shaped on deployment/governance.

Convergences and contradictions

Synthesis for RDCO

The market has quietly voted, and it voted RDCO's shape. RDCO's positioning has always been a single, accountable, centrally-owned agent-as-role — the COO agent, whose durable value lives in the vault (state), skills (tools), and MAC-style permissions (authority), not in a per-person sidekick. In 2026 the three largest enterprise-productivity vendors shipped defaults that say the same thing: the governed, invested-in, compounding unit is a role that someone owns and administers centrally, and the personal pet is a thin, ungoverned convenience layer. Every's retraction was the operator anecdote; Agentforce's "digital workforce," Microsoft's Agent 365 control plane, and Google's Agentspace registry are the same conclusion reached independently by three vendors with everything to lose. RDCO no longer has to argue that the roles pattern is right — it can cite the enterprise defaults as evidence and spend its breath on the harder question of who operates the role.

But the same evidence tightens the wedge. If Agentforce, Copilot Studio, and Workspace Studio all ship role-shaped agents that are easy to build ("in minutes, no code"), then "we'll build you a role-shaped agent" is being commoditized to a subscription line item — exactly the Archetype-1 conclusion of [[research/2026-05-21-enterprise-ai-agent-deployment-paths]], now sharper because each vendor also ships its own control plane. RDCO cannot win by being a cheaper Agent Builder. The defensible position is the layer the vendors structurally won't build: substrate-portable state and cross-substrate operation. Every vendor control plane is designed to govern agents inside its own stack — Agent 365 governs Microsoft's estate, Agentspace governs Google's, Agentforce governs Salesforce's. None has any incentive to own the client's agent-state independently of the vendor, because portability cannibalizes lock-in. The fractional agent-deployer who owns the client's role-definitions, evals, and operating context across whichever control plane they're locked into is filling the one seat the incumbents left empty. The enterprise default validates the shape; the multiplicity of vendor control planes creates the portability gap RDCO occupies.

There is a Sanity Check angle here, and it is not a product roundup (which would violate the no-derivative rule and just restate three vendor blogs). The original re-frame is the revealed-preference argument: enterprise agent defaults are a vote, and three independent, deep-pocketed actors cast identical ballots before most of their own customers finished a first pilot. The interesting claim isn't "here's what shipped" — it's that the pet model lost the org-physics argument in the market before the debate even reached most companies, and the vendors' governance spend is the tell. The piece writes the mechanism (why a role someone owns beats a sidekick everyone customizes — continuity, compounding, accountability, the fact that you can't put a pet on an org chart or fire it) and uses the three defaults + Every's retraction as convergent, independent evidence for a claim about how AI actually installs in organizations. The hook is that the "personal AI assistant for everyone" narrative that dominated 2024–2025 marketing is quietly being unwound by the same companies that sold it — and the receipts are in the release notes, not the keynotes.

Open follow-ups

Related

Sources

Vault:

Web (Microsoft):

Web (Google):

Web (Salesforce — direct fetch 403-blocked; framing via search index):