DataRobot: Most Business Processes Need Specialist Agents, Not Autonomous Generalists

Governance is where agents get stuck

DataRobot opens with a scenario: the person who must sign off on a first autonomous agent, such as a head of audit, a CISO or a risk lead, asks who authorized an action. If nobody can answer cleanly, the agent stays in pilot.

The company points to its Unmet AI Needs Survey 2026 of 413 agentic AI stakeholders. On whether existing access controls and catalogs are adequate to govern agents, 49% said yes and 51% said no. Governance was the top challenge companies expect when scaling from one agent to 100 or more, cited by 61%. According to the survey, 94% of organizations hit Day 2 operational issues after deploying agentic AI.

DataRobot says the common workaround, a human approving every meaningful action, automates nothing and "adds a new interface to the same bottleneck."

Long-running is not the same as autonomous

The post separates two properties. A long-running agent never turns off and keeps state across sessions. An autonomous agent can decide to act without a human approving that specific action. A dedicated flow such as a RAG pipeline or a ticket classifier is usually neither. Hermes-style agents are both.

DataRobot says a standing responsibility, such as month-end close, can justify both. The agent would pull reconciliations, flag exceptions, chase owners and escalate only what needs a person. The trade, the company says, is giving up a small, auditable surface for a bigger one that can do more unattended. It calls that worthwhile for judgment-heavy, long-horizon work, but usually not for a well-defined, high-volume, cost-sensitive task that a guarded flow handles at a fraction of the token spend.

The company notes that this agent class has gone mainstream. Open-source projects like Hermes Agent popularized it, and it says xAI shipped Grok Bot in August, Meta shipped Muse in September, and OpenAI launched Dots on September 29. According to DataRobot, Grok Bot and Muse each give the agent its own cloud machine, and Dots connects to more than 4,000 apps.

Why generalists are a shaky fit for business processes

DataRobot calls these products generalists, "the office administrators of the agent world." It sees them as a fine trade for a person delegating errands but a weaker one for a business process. Its reasons:

• A generalist reaches for a frontier model because it cannot know which step is hard, so work a small model could do runs on the most expensive one, on a loop.
• Broad access means a broad blast radius. DataRobot says Grok Bot itself calls its shared machine, logins and files "a real blast radius."
• Consumer-first products come with consumer-first defaults on data use and regional availability, which is a different bar from what a regulator expects.
• Their guardrails lean on per-user approvals, while enterprises need policy that applies to every agent and every action.

What autonomy asks of security teams

The post lists risks that a stateless pipeline never posed. Long-lived credentials become a permanent target. Prompt injection turns into privilege escalation: a poisoned document gives a pipeline a bad answer, but a long-running agent could open a PR, leak a secret or message a customer. Memory becomes a durable attack surface, and because behavior drifts as memory and skills evolve, a one-time code review is not enough and tracing must be always on.

Identity is the unsolved part, DataRobot says. A team agent is provisioned once and used by many, permissions tend to ratchet up, sub-agents can act without a human originator, and a shared service identity erases per-user attribution. Its recommended posture resembles treating a highly privileged contractor: a named identity where possible, scoped and reviewed grants, short-lived tokens, and an audit trail separating "the agent did this" from "the agent did this because a specific person asked."

DataRobot's answer: Decision Agents

For truly open-ended problems, DataRobot says to run a self-learning agent inside your own governed perimeter, with tracing from day one. For most critical business problems, it recommends a specialist that executes the process, with governance and approval gates built in.

Its product for this is Decision Agents, each built to make and carry out the decisions behind a specific business process. According to the company, governance comes from the Agent Workforce Platform they run on. Through the DataRobot LLM gateway, users choose the model per task. Guard models check key steps, and human approval gates sit where owners decide. Each agent gets its own identity and a named owner, secrets live in the DataRobot credentials store, and OpenTelemetry traces of model and tool calls give audit teams a record. Agents can run in the cloud, on-prem or air-gapped.

The pitch for buyers is predictable outcomes and plannable spend without the standing risk surface of a long-running generalist. The post also mentions DataRobot Labs, which has published experimental open-source projects for the open-ended end of the spectrum. The source text is cut off before describing them.