agentic_ai_basics

agent

Agent — an AI system that autonomously perceives, reasons, plans, and acts to achieve a goal, beyond simple question-answering.

agent.py

A budget-checking agent. Give it one goal (e.g. “do these purchases fit under my budget?”) and pursue_goal() runs unattended — deciding its own steps, tool calls, and when it’s done — until it produces a final answer or hits MAX_STEPS.

Concepts covered

• Autonomy over the whole task, not just one call — contrast with ../basics/basic.py’s ask_claude: one message in, one message out, no notion of “working on it.” Here, a single call to pursue_goal() runs an entire perceive → reason → plan → act cycle, repeatedly, deciding on its own how many steps the goal needs.
• No human between steps — contrast with ../tool_use/basic_agentic_tools.py’s run_turn(), which waits for a fresh human message before every action. This agent gets ONE goal up front and keeps going without further input until it’s satisfied or capped out.
• The agent decides when it’s done — nothing in this script tells the agent the goal is met. response.stop_reason != "tool_use" is checked after every call; the agent itself chooses to stop requesting tools once it judges it has enough.
• Visible planning via record_finding — a tool with no real computation behind it, whose only purpose is to make the agent’s intermediate reasoning show up as tool calls in the transcript rather than staying implicit.
• MAX_STEPS as an unconditional backstop — same idea as ../../Execution_Loops/max_iterations/max_iterations.py: autonomy without a hard cap is dangerous, since a confused agent could otherwise loop on tool calls indefinitely.

Run

From the repo root:

pip install -r requirements.txt
export ANTHROPIC_API_KEY=your-key-here
python3 Core_Architecture/agent/agent.py

Try (or just press enter for the built-in default):

Goal: I want to buy 3 items priced at $12.50, $7.25, and $19.99. My budget is $50. Do they fit, and how much would I have left over?

[step 1] act: calculate({'expression': '12.50 + 7.25 + 19.99'})
[step 1] perceive: 12.50 + 7.25 + 19.99 = 39.74

[step 2] act: record_finding({'finding': 'total is 39.74, under the $50 budget'})
[step 2] perceive: Noted: total is 39.74, under the $50 budget

[step 3] agent finished on its own

Final answer: Yes, they fit under your $50 budget, with $10.26 left over.

Configuration

• MODEL, MAX_TOKENS, EFFORT, SYSTEM_PROMPT — see ../basics/README.md for the first three
• MAX_STEPS — the hard cap on perceive-reason-act cycles (default: 8)
• TOOLS / execute_tool — the calculate and record_finding tools available to the agent while it works

See also

• ../basics/README.md — the single request/response call this template’s autonomy is contrasted against
• ../tool_use/README.md — the underlying tool-calling loop mechanics (TOOLS, execute_tool, parallel calls), here repurposed for unattended goal pursuit instead of turn-by-turn chat
• ../llm_backbone/README.md — swapping which model powers this same kind of loop
• ../../Execution_Loops/max_iterations/README.md — the unconditional-cap pattern MAX_STEPS borrows