agentic_ai_basics

agentic_loop

Agentic Loop — the core observe → think → act → observe cycle an agent runs until the task is done or halted, in its simplest possible form: a multi-turn chat built on top of a stateless API.

basic_agentic_loop.py

A continuous chat agent. Keeps asking you questions and answering on any topic until you type exit. Maintains full conversation history across turns.

Concepts covered

The API is stateless — Claude has no memory between calls. Every request is independent, so your code must track everything said so far and resend the full history each time.
The messages list as the conversation — each entry is one turn (role + content); it starts empty and grows by two entries (user, then assistant) every loop iteration.
The agentic loop pattern — while True: get input, call the model with accumulated state, act on the result, repeat until an exit condition. This is the shape behind every conversational agent, tool-using or not.

Mapping the loop onto observe → think → act → observe

This template’s loop is the cycle at its most stripped-down — worth naming explicitly, since every other template in Execution_Loops/ and most of the rest of the repo is a variation on it:

Phase	In this template
Observe	`input("You: ")` — read what the user just said
Think	the API call itself — Claude reasons over the accumulated `messages` history
Act	`print(f"\nClaude: {reply}\n")` — the only “action” here is responding in text; there’s no tool call, so nothing actually changes in the world
Observe (again)	back to `input("You: ")` for the next cycle

The honest limitation: because this loop never calls a tool, “Act” never touches anything external, and there’s nothing for a subsequent “Observe” to notice besides the user’s next message — the cycle only closes through the human, not through the environment. For the version where Act means calling a tool and Observe means reading back that tool’s real result — the fuller form of the cycle — see ../../Core_Architecture/tool_use/basic_agentic_tools.py.

Run

From the repo root:

pip install -r requirements.txt
export ANTHROPIC_API_KEY=your-key-here
python3 Execution_Loops/agentic_loop/basic_agentic_loop.py

You’ll see:

Chat with Claude. Type 'exit' to end the conversation.

You:

Type a message and press Enter to get a response. Keep chatting, or type exit to end the conversation.

Configuration

Edit the constants at the top of basic_agentic_loop.py:

MODEL — the Claude model used (default: claude-sonnet-5)
MAX_TOKENS — max response length
EFFORT — thinking/response depth (default: medium)
SYSTEM_PROMPT — instructions that shape Claude’s behavior as the assistant