The next fight in enterprise AI is over the layer that governs agents
The discussion about artificial intelligence agents has entered a new stage. After testing assistants to summarize documents and answer questions, companies are starting to evaluate systems capable of consulting different sources and carrying out tasks within internal processes.
This movement shifts the center of the decision. The question is no longer just which model offers the best answer. It becomes how the company intends to let AI operate.
In September, OpenAI presented the Agents API, infrastructure for building and running long-running agents, with context management, tool use, subagents and execution environments. A few days later, Salesforce announced the Enterprise AI Harness concept, an architecture that brings together context, action, governance, security and models in a common control layer.
The announcements point in the same direction: agents need a structure that connects them to the business.
The model is just one part of the system
To answer whether an order can be delivered today, an agent may need to consult the customer record, inventory, contract, commercial terms and logistics lead times. None of these sources, on its own, holds the complete answer.
The challenge is organizing this information without granting unrestricted access to the company's systems. This layer includes data, processes, definitions, history and internal rules. Models can change quickly; organized knowledge about the operation remains particular to each business.
The meaning of the data also needs to be aligned. “Active customer”, “available order” or “minimum margin” may represent different criteria across areas. Without these definitions, the agent can gather correct information and still reach an inadequate conclusion.
Autonomy needs limits
Before releasing an agent, the organization needs to define which systems can be consulted, which tasks can be carried out and when human approval will be required. Actions should also be logged, with mechanisms to stop the flow when something goes off track.
An agent can produce a recommendation in a controlled environment, while a change in a financial, commercial or operational system requires additional validation. This care is especially important when there is confidential data, commitments to customers or decisions that are hard to undo.
The strategic decision, therefore, doesn't start with choosing the tool. It starts with defining how AI should take part in the work. The edge won't be in having more agents, but in connecting them to the right context and letting them act within understandable limits.
Sources
OpenAI, Sep 2026. https://openai.com/index/introducing-the-agents-api/
Salesforce, 11 Sep 2026. https://www.salesforce.com/ap/news/press-releases/2026/09/11/salesforce-introduces-the-trusted-enterprise-ai-harness/