AI Agent

An AI agent is a system built around an LLM that can autonomously reason about tasks, plan sequences of actions, execute those actions using external tools, observe the results, and iterate until a goal is achieved.

Agent architecture:

1. Perception: receive a task description and observe current state (files, terminal output, API responses)
2. Reasoning: LLM analyzes the situation and decides the next action
3. Action: execute a tool call (run a shell command, edit a file, call an API, search the web)
4. Observation: receive the result of the action
5. Loop: repeat steps 2-4 until the task is complete or the agent determines it cannot proceed

Core capabilities that distinguish agents from chatbots:

• Tool use: call external functions (file I/O, shell commands, APIs, databases)
• Multi-step planning: break a complex goal into sequential sub-tasks
• State management: maintain context across many actions within a single task
• Error recovery: detect failures and try alternative approaches
• Autonomy levels: from “ask before every action” to “complete the task and report back”

Agent frameworks and patterns:

• ReAct (Reasoning + Acting): interleave reasoning traces with tool calls
• Tool-augmented generation: LLM selects and invokes tools from a defined toolkit
• Multi-agent systems: multiple specialized agents collaborate on complex tasks
• Human-in-the-loop: agent pauses for human approval on risky actions

Safety considerations:

• Authority boundaries: limit what actions the agent can take (read-only vs. write access)
• Guardrails: prevent harmful actions (deleting production data, exposing secrets)
• Audit trails: log every action for review
• Sandboxing: run agent actions in isolated environments

# Simplified agent loop
import anthropic

client = anthropic.Anthropic()
tools = [
    {
        "name": "run_command",
        "description": "Execute a shell command",
        "input_schema": {
            "type": "object",
            "properties": {
                "command": {"type": "string", "description": "The command to run"}
            },
            "required": ["command"]
        }
    }
]

messages = [{"role": "user", "content": "Check disk usage and find the largest directory"}]

# Agent loop: reason, act, observe, repeat
while True:
    response = client.messages.create(
        model="claude-sonnet-4-20250514",
        max_tokens=1024,
        tools=tools,
        messages=messages,
    )

    if response.stop_reason == "end_turn":
        print("Agent finished:", response.content[0].text)
        break

    if response.stop_reason == "tool_use":
        tool_call = next(b for b in response.content if b.type == "tool_use")
        result = execute_command(tool_call.input["command"])
        messages.append({"role": "assistant", "content": response.content})
        messages.append({"role": "user", "content": [
            {"type": "tool_result", "tool_use_id": tool_call.id, "content": result}
        ]})