AI Agents With Tools

The agent idea, returned

Modern AI agents wrap a language model in a loop: it reads the goal and context, decides on an action (call a tool such as search, a calculator, a database or code execution), observes the result, and continues until done. This brings back the classic agent framework from the start of this course, with an LLM as the decision-maker. Agents can automate multi-step tasks, but they also compound errors, can be manipulated by malicious content (prompt injection) and need limits: narrow tool permissions, approval for risky actions, step budgets and evaluation.

An agent loop (sketch)

The classic perceive-decide-act cycle with an LLM. Not run here.

state = {"goal": user_goal, "observations": []}
for step in range(MAX_STEPS):                      # budget: never loop forever
    action = llm_decide(state)                     # e.g. {"tool": "search", "args": {...}} or {"final": "..."}
    if "final" in action:
        return action["final"]
    if needs_approval(action):
        action = ask_human(action)                 # risky actions need a person
    result = run_tool(action)                      # tools with least privilege
    state["observations"].append(result)
return "stopped: step budget reached"

Budget every loop

Limit steps, tokens and time per task; an agent without limits can run up costs or repeat mistakes.

Quick check: What do modern AI agents add to a language model?

• Removal of all safety limits
• A larger font
• A faster keyboard
• A loop of deciding, calling tools and observing results