Agents as Organizational Actors

A year ago, most companies talked about artificial intelligence agents like chatbots with better memory. In 2026, vendors are selling them more like junior staff that can search the web, open files, click through software, and finish multi-step work with a trace of what happened. (openai.com) That shift is visible in the product language itself. OpenAI’s Agents software development kit says agents “plan, call tools, collaborate across specialists, and keep enough state to complete multi-step work,” which is much closer to a workflow role than a help box. (developers.openai.com) Once software can act instead of just answer, the first question stops being “is the model smart enough.” The first question becomes who is allowed to let it touch a browser, a customer record, or a payment system. (openai.com) Anthropic made the same turn when it described computer use in October 2024. Its system could move a cursor, click on screen locations, and type through a virtual keyboard, which means the model was no longer limited to text boxes and application programming interfaces. (anthropic.com) By March 23, 2026, Anthropic had pushed that capability into a research preview for Claude users on Pro and Max plans. The release notes said Claude could open files, run developer tools, point, click, and navigate what was on a user’s screen “with no setup required.” (support.claude.com) That is why audit trails suddenly moved from a compliance footnote to a buying requirement. If an agent can click through a live system, a company needs the software equivalent of a security camera that records who told it what, which tools it used, and what it changed. (developers.openai.com) OpenAI now pitches “full trace” as a built-in feature of its Agents software development kit. The company’s agent platform also advertises versioning and guardrails, which are management features you add when the software is expected to do work repeatedly inside an organization. (developers.openai.com) (openai.com) Microsoft’s pricing pages show the same change from seat counting to activity counting. Copilot Studio tells customers to forecast “Copilot Credits” based on agent type, traffic, orchestration, knowledge, and tools, which treats an agent more like a metered worker than a licensed employee. (learn.microsoft.com) Salesforce is moving in the same direction. Its Agentforce pricing page offers consumption-based pricing through Flex Credits or Conversations alongside per-user licensing, which is what you do when some of the “users” are software acting on behalf of humans. (salesforce.com) That pricing fight sounds technical, but it lands in procurement first. A company that used to buy 5,000 software seats now has to ask whether 500 people plus 2,000 agent-run tasks is cheaper, riskier, or harder to control than the old model. (learn.microsoft.com) (salesforce.com) So the real story is not that agents got a little better at answering prompts. The real story is that major vendors now package them as operational actors with tools, traces, guardrails, and usage meters, which forces companies to decide how much authority software should have before it starts looking less like an assistant and more like a member of the org chart. (openai.com 1) (openai.com 2)