Agentic AI in Claude

An agent differs from a chat assistant that answers once and then stops. Claude works as an agent by cycling through three phases — gather context, take action, and verify results — adapting the loop to the task and continuing until the goal is met, rather than stopping after a single response.

Acting rather than merely responding involves using tools such as web search, file operations, and code execution, rather than relying solely on what Claude learned during training.

Claude requests a tool, the tool runs and returns a result, which Claude uses to decide its next step. A useful result can move the task forward, while an unhelpful result, such as an empty search, can prompt Claude to try a different approach.

The same agent loop and tool-calling approach applies across all Claude surfaces, even though their capabilities and use cases differ. What changes is the environment, the available tools, and who each surface is designed for — not the underlying agentic pattern itself.

Claude Code tackles a software engineering task by first exploring the relevant parts of a codebase to gather context. It uses development tools to make changes, then verifies the result by running tests. If a test fails, Claude uses that output to revise the fix and retry rather than stopping after a single attempt.

Claude in Chrome completes multi-step browser tasks by reading a webpage to gather context, then navigating and interacting with it as a person would. It is subject to permission and safety controls and pauses for confirmation before taking higher-risk actions on the user’s behalf.

Claude Cowork performs complex knowledge work for people who don’t need a development environment. It gathers context from connected folders and apps, uses information from one step to inform the next, and delivers a finished output rather than leaving someone to assemble the results manually.

The Claude Agent SDK is not a finished product; it is a developer toolkit. It provides components for model calls, tool execution, and context management, enabling developers to build custom agents that can plan, use tools, and execute multi-step tasks within their applications.

Connecting an MCP server gives Claude access to external tools, data sources, and services beyond its built-in capabilities. MCP defines a client that runs in the host application and a server that provides access to the external system. Once connected, Claude can call the server’s tools and use their results to continue the task.

The appropriate level of oversight for an agentic task depends on risk, reversibility, and consequences. Requiring approval for every action adds friction, while requiring none risks an unwanted or unrecoverable outcome. The goal is the right balance between autonomy and oversight for each action, not the maximum of either.

Delegation means Claude hands off parts of a complex task to sub-agents that work in parallel, then combines their results. It is useful when those parts can be completed independently; if one step depends on another’s result, delegation adds coordination with little benefit.

A strong instruction for an agentic task defines the goal, relevant context, constraints or boundaries, available resources such as files, tools, or connectors, success criteria, and the expected level of autonomy. Success criteria matter most because they allow Claude to verify its work against the goal before stopping.

The right Claude surface for a task depends on where and how the work needs to happen, not just the task description. Multi-step work that produces a deliverable fits Claude Cowork; work on a codebase fits Claude Code; work on a live webpage fits Claude in Chrome; and building agent behavior into another application fits the Agent SDK.