Traditional LLMs follow a single prompt → single response pattern. AI agents have an autonomous execution loop: given a goal, they plan, call tools, verify results, and retry if needed.
Aspect
LLM Prompting
AI Agent
Execution
Single request-response
Multi-step loop (observe → think → act → repeat)
External integration
Limited
Tool calling (APIs, DBs, filesystem)
Autonomy
User directs each step
Given a goal, decomposes and executes independently
When agents are unnecessary: Simple Q&A, well-defined pipelines (Step Functions, etc.), real-time response requirements.
LLM chat was confined to the browser. Desktop Agents break this limit with local file access, OS manipulation (Computer Use), external tool connection (MCP), and long-running autonomous execution.
Claude Cowork status (2026.09): macOS/Windows GA (April) → web, iOS, Android + cloud remote sessions (July) → built-in browser in the desktop app side panel (announced late August, rolling out to Pro/Max/Team). Cross-device session continuity and Artifacts sharing.
Through 2026, the major vendors all moved in the same direction — from chatbots that answer requests to always-on, proactive agents that act on their own. Here, “always-on” means the agent keeps running in the background even after you close the app, and “proactive” means it suggests and performs work first, without waiting for a command.
These products share four traits.
Background execution — work continues even when you step away.
Dedicated runtime — each agent runs on its own cloud computer or VM.
Proactive checks — it periodically checks connected apps (mail, calendar, etc.) and surfaces suggestions first.
Control over sensitive actions — it supports human approval and an audit trail for actions like sending, paying, or deleting (whether approval is required depends on the product and the organization’s settings).
The representative product from each vendor is below. “Category” indicates consumer vs. enterprise.
New “Autopilots” category. Per-agent Entra identity, credentials redacted from logs, Purview DLP enforced in the moment. Approval for sensitive actions follows the organization’s settings
As agents and tools multiply in an organization, teams rebuild the same capabilities and it gets hard to track who built what. Two layers address this — a registry is the catalog for finding what exists, while a gateway is the runtime entry point that tool calls actually pass through. They are complementary, not alternatives.
Major vendors offer registry/directory services, and many extend their scope beyond their own platform to third-party and on-premises agents. The emphasis differs by vendor.
Determined whether an agent is actually needed (vs. simple prompting)
Least-privilege + allowlist per tool
Guardrails (input/output/execution limits)
Human-in-the-Loop policy defined
Tracing & monitoring (OpenTelemetry)
Cost budgets and circuit breakers
When operating many agents: inventory (registry) of agents, tools, and MCP servers plus a publish-approval process — apply execution control separately via gateway and runtime policies