Module 05: AI Agents and Tool Calling

Goal

Build agents that use tools safely, validate actions, preserve state, support human confirmation, and create audit logs.

Core Topics

• Tool calling.
• Function calling.
• Agent loops.
• Planning.
• Memory and state.
• Workflows.
• Human-in-the-loop.
• Multi-agent systems.
• Model Context Protocol.
• Agent evaluation.
• Agent security.

Basic Agent Flow

User -> LLM -> Tool Selection -> Tool Execution -> Observation -> LLM -> Final Response

Production Agent Flow

User -> Authentication -> Authorization -> Intent -> Policy Check -> Agent -> Tool -> Validation -> Human Confirmation -> Execution -> Audit Log

Study Tasks

• Build a calculator tool.
• Build a database read tool.
• Build a write tool that requires confirmation.
• Add schema validation for tool inputs and outputs.
• Add role-based tool authorization.
• Add audit logs.
• Add agent evaluation cases.
• Add retry and failure handling.

Project

Build projects/05-erp-ai-agent/.

Safety Rules

• Read operations may execute after validation and authorization.
• Write operations require validation, preview, user confirmation, idempotency, execution, and audit.
• Destructive operations require stronger confirmation and rollback planning.
• Tool permissions must not rely on model behavior alone.

Tips

• Prefer deterministic workflows when the process is predictable.
• Use agents when flexible planning and tool selection are truly needed.
• Keep tools narrow and well-typed.
• Test malicious or ambiguous user requests.

Completion Checklist

• Can explain agent vs workflow trade-offs.
• Can implement tool calling.
• Can secure tool execution.
• Can require confirmation for writes.
• Can evaluate agent task success.