Enterprise AI agent rush
The race in AI is moving from raw models to full‑service agents that run inside business workflows, not just answer questions. Anthropic this week launched Claude Managed Agents to handle the infrastructure, orchestration and governance enterprises need to deploy autonomous workflows, and other vendors are framing their strategies the same way as they chase integration and reliability over benchmark scores. That shift matters because buyers will judge vendors on deployment, observability and permissioning rather than model IQ alone, and Anthropic says its ARR has jumped to roughly $30bn annualised — a sign enterprise spend is consolidating quickly. (wired.com (startuphub.ai)
The new fight in artificial intelligence is over who can run the boring, risky middle of office work: logging into systems, carrying state from one step to the next, and not breaking the rules while doing it. Anthropic pushed straight at that problem on April 8 with Claude Managed Agents, a public beta service for cloud-hosted agents that it says can cut production work from months to days. (claude.com) Anthropic’s pitch is not “our model is smarter.” Its pitch is that companies should stop wiring up sandboxed code execution, checkpointing, credential management, scoped permissions, tracing, and error recovery by hand every time they want an agent to do a job. (claude.com) That sounds dry until you picture the difference between a chatbot and an employee. A chatbot answers one question in one window; an agent has to remember what happened three steps ago, call tools in the right order, recover when one tool fails, and leave a record that a compliance team can inspect later. (opensource.microsoft.com) Microsoft is describing the same shift from a different angle. In its April 2 Agent Governance Toolkit launch, Microsoft said agents are already booking flights, executing trades, writing code, and managing infrastructure, and it framed the next bottleneck as runtime security rather than model quality. (opensource.microsoft.com) IBM is selling the same missing layer in corporate language: observability and governance. Its watsonx Orchestrate materials say companies now want dashboards for latency, tool-call reliability, completion rates, policy controls, and alerts across systems like Salesforce, SAP, Workday, Amazon Web Services, and Microsoft 365. (ibm.com) Microsoft’s own Azure Foundry Agent Service uses nearly identical words. Azure says the product is for securely designing, deploying, and scaling agents, with governance, observability, and more than 1,400 action connectors through Azure Logic Apps. (azure.microsoft.com) Google Cloud is also talking less about one brilliant model and more about connected workflows. In its 2026 agent trends report, Google says businesses are moving from one-off prompts to “digital assembly lines” that run whole processes, and it points to cross-platform agents built with Salesforce using the Agent2Agent protocol. (cloud.google.com) (blog.google) OpenAI has been moving in the same direction too. Its Operator product was introduced as a computer-using agent, and OpenAI said it planned to expand those capabilities to Team and Enterprise users and fold them into ChatGPT once safety and usability held up at scale. (openai.com) The reason this is happening now is that companies learned the hard part of “agentic” software is not getting a model to sound capable in a demo. The hard part is letting software touch payroll systems, customer records, source code, procurement tools, and regulated workflows without losing track of who approved what. (anthropic.com) (opensource.microsoft.com) Anthropic has been explicit about that regulated-industry gap. In its February partnership with Infosys, it said the joint work would focus on telecommunications, financial services, manufacturing, and software development, with governance and transparency built in for industries that need precision and compliance. (anthropic.com) That is why the buying checklist is changing. A chief information officer comparing vendors in 2026 is now looking at permissioning, tracing, connectors, policy enforcement, and rollback controls in the same way cloud buyers once compared uptime and security certifications. (claude.com) (ibm.com) (azure.microsoft.com) Anthropic’s revenue number shows how quickly that enterprise money is bunching up. Multiple reports this week said Anthropic disclosed a run-rate revenue figure above $30 billion, up from about $9 billion at the end of 2025, which suggests large companies are already picking platforms and spending heavily once the tooling looks safe enough to deploy. (economictimes.indiatimes.com) (fortuneindia.com)
