AI Agents

Building AI agents with tools and planning.

Agent Architecture

┌─────────────────────────────────────┐
│            AI AGENT                  │
├─────────────────────────────────────┤
│  ┌──────────┐                       │
│  │  BRAIN   │  (Foundation Model)   │
│  │ Planning │                       │
│  │ Reasoning│                       │
│  └────┬─────┘                       │
│       │                             │
│   ┌───┴───┐                         │
│   ↓       ↓                         │
│ ┌─────┐ ┌──────┐                    │
│ │TOOLS│ │MEMORY│                    │
│ └─────┘ └──────┘                    │
└─────────────────────────────────────┘

Tool Definition

tools = [{
    "type": "function",
    "function": {
        "name": "search_database",
        "description": "Search products by query",
        "parameters": {
            "type": "object",
            "properties": {
                "query": {"type": "string"},
                "category": {"type": "string", "enum": ["electronics", "clothing"]},
                "max_price": {"type": "number"}
            },
            "required": ["query"]
        }
    }
}]

Planning Strategies

ReAct (Reasoning + Acting)

REACT_PROMPT = """Tools: {tools}

Format:
Thought: [reasoning]
Action: [tool_name]
Action Input: [JSON]
Observation: [result]
... repeat ...
Final Answer: [answer]

Question: {question}
Thought:"""

def react_agent(question, tools, max_steps=10):
    prompt = REACT_PROMPT.format(...)

    for _ in range(max_steps):
        response = llm.generate(prompt)

        if "Final Answer:" in response:
            return extract_answer(response)

        action, input = parse_action(response)
        observation = execute(tools[action], input)
        prompt += f"\nObservation: {observation}\nThought:"

Plan-and-Execute

def plan_and_execute(task, tools):
    # Step 1: Create plan
    plan = llm.generate(f"Create step-by-step plan for: {task}")
    steps = parse_plan(plan)

    # Step 2: Execute each step
    results = []
    for step in steps:
        result = execute_step(step, tools)
        results.append(result)

    # Step 3: Synthesize
    return synthesize(task, results)

Reflexion (Self-Reflection)

def reflexion_agent(task, max_attempts=3):
    memory = []

    for attempt in range(max_attempts):
        solution = generate(task, memory)
        success, feedback = evaluate(solution)

        if success:
            return solution

        reflection = reflect(task, solution, feedback)
        memory.append({"solution": solution, "reflection": reflection})

Memory Systems

class AgentMemory:
    def __init__(self):
        self.short_term = []        # Recent turns
        self.long_term = VectorDB() # Persistent

    def add(self, message):
        self.short_term.append(message)
        if len(self.short_term) > 20:
            self.consolidate()

    def consolidate(self):
        summary = summarize(self.short_term[:10])
        self.long_term.add(summary)
        self.short_term = self.short_term[10:]

    def retrieve(self, query, k=5):
        return {
            "recent": self.short_term[-5:],
            "relevant": self.long_term.search(query, k),
        }

Agent Evaluation

def evaluate_agent(agent, test_cases):
    return {
        "task_success": mean([agent.run(c["task"]) == c["expected"] for c in test_cases]),
        "avg_steps": mean([agent.step_count for _ in test_cases]),
        "avg_latency": mean([measure_time(agent.run, c["task"]) for c in test_cases]),
    }

Best Practices

1. Start with simple tools, add complexity gradually
2. Add reflection for complex tasks
3. Limit max steps to prevent infinite loops
4. Log all agent actions for debugging
5. Use evaluation to measure progress