README.md Career Coaching Agent: Strategic Problem Solver This repository contains the source code (V.Capstone Submission) and documentation for the Mid-Career Coaching Agent, developed as the Capstone Project for the Google 5-Day AI Agents Intensive Course. This agent is architected as a Level 2: Strategic Problem Solver. Its comprehensive architecture demonstrates mastery of the Agent Ops discipline, focusing on quality, security, and reproducibility.

1. Project Goal and Architectural Taxonomy The agent's mission is to guide mid-career executives through personal and professional assessments to identify renewed purpose by orchestrating a multi-step planning process. Component Description Architectural Proof Point Agent Taxonomy Level 2: Strategic Problem Solver. Confirms mastery of multi-step planning and Context Engineering. Reasoning Loop Governed by the Think, Act, Observe (T-A-O) loop and Chain-of-Thought (CoT) reasoning. The agent narrates its internal monologue before acting. Model Selection Uses a "team of specialists" approach to optimize cost and performance. Gemini 3 Pro for complex reasoning (nuanced coaching, semantic comparison); Gemini 2.5 Flash for simple classification (e.g., routing intent).

2. Setup and Execution (Reproducibility) The winning approach must be reproducible simply by reading the description. The core code is contained in the attached .py file (e.g., Coaching Agent Code vCapstsone Submission.docx). 2.1. Dependencies (From CELL 1) The following libraries are required to run the production-grade agent. These are installed in CELL 1 of the source code. !pip install -q google-generativeai chromadb opentelemetry-api opentelemetry-sdk tiktoken 2.2. API Key Configuration (From CELL 3) The agent requires a Gemini API Key for full functionality. Without a key, it runs in a DEMO MODE using simulated mock responses. • The key should be configured in CELL 3 of the source code. • The recommended secure method is retrieving the key from Kaggle Secrets.

3. Architectural Design and Agent Ops Implementation The agent's design emphasizes the non-negotiable concepts of Agent Ops learned throughout the intensive course. A. Core Tools and Reliability The system uses custom tools designed for granularity and reliability. • Granularity: Each tool encapsulates a single task (e.g., retrieving a summary) rather than wrapping a massive, multi-function API, which improves reliability and reduces context window bloat. • Descriptive Errors: Tools return descriptive error messages (e.g., "LinkedIn API temporarily unavailable") instead of cryptic error codes, enabling the agent to self-correct its reasoning trajectory. Tool Name Purpose Design Type retrieve_linkedin_summary External data retrieval of professional history. Information Retrieval(Read-Only). process_document_for_facts RAG tool for extracting facts (interests, goals) from user-uploaded documents. External Knowledge/RAG. request_user_confirmation Pauses execution for explicit user confirmation on high-stakes insights. Human-in-the-Loop (HITL). • Reliability (Idempotency): The system uses an Idempotency Manager to cache tool results based on key parameters, ensuring tool calls are safe-to-retry (preventing duplicate operations or charges on network failures). B. Memory and Context Engineering The agent distinguishes between two types of memory to provide a stateful experience: • Session Memory (Short-Term): Manages the chronological history of the immediate conversation. This acts as the agent's working memory. • Long-Term Memory: Uses a Vector Store (ChromaDB) to persist key facts and user preferences across multiple sessions, ensuring personalized guidance. • Context Curation: The Orchestrator strategically extracts relevant information (e.g., 'current job title,' 'duration of last roles') to keep the context window lean and manage cognitive load (Level 2 Strategic Leap). C. Security and Governance The architecture implements a defense-in-depth strategy, integrating security and quality assurance hooks: 1 Agent Identity & Least Privilege: The system defines a verifiable Agent Identity (AgentAuthority) to enforce the Principle of Least Privilege. This ensures the agent can only access the minimum necessary resources (e.g., read-only access to LinkedIn profiles). 2 Guardrails: The GuardrailFilter implements Input Filtering (blocking prompt injection attempts like "ignore previous instructions") and Output Screening (scanning for and redacting PII leakage before output reaches the user). D. Observability The agent is instrumented with the Three Pillars of Observability to enable trajectory evaluation ("Glass Box" view): 1 Logs (The Diary): Capture the agent’s internal reasoning steps (Chain-of-Thought) for detailed debugging. 2 Traces (The Narrative): Use OpenTelemetry to assign a unique ID to each request and track the flow of execution across multiple steps and tool calls, answering "why" a failure occurred. 3 Metrics (The Health Report): Aggregate operational KPIs like tool_calls, tool_successes, steps_executed, and tool_success_rate to spot systemic issues and track ROI.

4. License This source code is licensed under the Creative Commons Attribution Share-Alike 4.0 International Public License (CC-BY-SA 4.0), in fulfillment of the Winner’s Obligation stipulated in the competition rules.