Agentic AI

Anthropic function calling — tool use loop

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();

const tools = [
  {
    name: "execute_code",
    description: "Run Python code in a sandbox and return stdout/stderr",
    input_schema: {
      type: "object",
      properties: {
        code: { type: "string", description: "Python code to execute" }
      },
      required: ["code"]
    }
  },
  {
    name: "web_search",
    description: "Search the web and return top results",
    input_schema: {
      type: "object",
      properties: {
        query: { type: "string" }
      },
      required: ["query"]
    }
  }
];

// Agentic loop
let messages = [{ role: "user", content: "Scrape the top 10 HN stories and plot upvotes" }];
while (true) {
  const response = await client.messages.create({
    model: "claude-opus-4-5", max_tokens: 4096, tools, messages
  });
  
  if (response.stop_reason === "end_turn") break;  // done
  
  // Process tool calls
  for (const block of response.content) {
    if (block.type === "tool_use") {
      const result = await executeTool(block.name, block.input);
      messages.push({ role: "assistant", content: response.content });
      messages.push({ role: "user", content: [{ type: "tool_result", tool_use_id: block.id, content: result }] });
    }
  }
}

1. Plan: "I should load the CSV, examine its structure and data types, then create an appropriate visualization."
2. Act: Calls execute_code with Python to load the CSV and print df.head(), df.dtypes, df.describe().
3. Observe: Sees columns [date, revenue, users, region], revenue and users are numeric, region is categorical.
4. Re-plan: "A grouped bar chart of revenue by region over time would show the key trends."
5. Act: Calls execute_code with matplotlib/seaborn code to generate the chart.
6. Observe: Code raises a ValueError — date column is a string, not datetime.
7. Self-correct: Adds pd.to_datetime(df["date"]) to the code, re-executes.
8. Observe: Chart generated successfully at /tmp/chart.png.
9. Complete: Returns the chart + written interpretation of the trends it shows.

1. What is the ReAct (Reasoning + Acting) pattern in agentic AI? (Answer: ReAct (Yao et al. 2022): the agent alternates between Reasoning steps (internal thought about what to do next) and Acting steps (calling a tool). Format: Thought: I need to find the current stock price... Action: search[AAPL stock price today] Observation: AAPL is at $189.32 Thought: Now I can answer... Final answer: AAPL is $189.32. The explicit reasoning steps make the agent's decisions interpretable and allow self-correction if a reasoning step is wrong.)
2. What are the main failure modes of AI agents? (Answer: (1) Hallucination of tool results: agent 'recalls' information instead of calling the tool, fabricating results. (2) Infinite loops: agent calls the same tool repeatedly without progress. (3) Prompt injection from tool results: malicious content in search/document results hijacks agent behaviour. (4) Over-permissive tool use: agent takes irreversible actions (send email, delete file) without confirming. (5) Context overflow: long agentic runs accumulate context until the window fills, causing the agent to lose track of its goal.)
3. What is the difference between a single-agent and multi-agent system? (Answer: Single-agent: one LLM with multiple tools, handles all subtasks sequentially. Simpler, easier to debug. Multi-agent: multiple specialised LLM agents, each with specific tools and expertise — an orchestrator coordinates them. Example: researcher agent (web search) + analyst agent (code execution) + writer agent (drafting). Advantages: specialisation (each agent optimised for its role), parallelism (agents work concurrently), scalability. Disadvantages: coordination overhead, error propagation between agents, debugging difficulty.)
4. How does human-in-the-loop oversight work in production AI agent systems? (Answer: Checkpoints: the agent pauses at predefined high-risk action points (sending email, making API calls with financial impact, deleting data) and requires explicit user approval before proceeding. Confirmation prompts: 'I plan to send this email to 500 contacts — confirm?' Audit logging: every tool call, result, and decision is logged for review. Rollback: reversible actions use transactions or staging environments before committing. Interrupt handlers: users can stop the agent mid-run. Minimal footprint principle: agents request only necessary permissions and prefer reversible actions.)
5. What is the difference between agentic AI and traditional RPA (Robotic Process Automation)? (Answer: RPA: rule-based automation of structured workflows — scrapes data from fixed UI positions, follows rigid decision trees. Brittle: breaks when UI changes. Cannot handle ambiguity or novel situations. Agentic AI: understands intent from natural language, adapts to UI changes, handles ambiguous situations through reasoning, and can decompose novel tasks it has never seen before. RPA is scripted; agents are reasoning. Hybrid approaches (AI + RPA) use agents for decision-making and RPA for reliable execution of structured steps.)