Day 2 (morning): Agentic Workflows in Code - CrewAI from One Agent to a DevOps Crew to a Parallel Investment Adviser with Custom Tools

Ishan's distinction: an agent re-invokes the model, judges the output, iterates, then returns. A raw LLM cannot read a file, browse or query a database - text in, text out - so a tool (web search, code interpreter, file reader, DB connector) is the mechanism that gives an agent a hand in the world. Memory splits into short-term (this execution) and long-term (all past executions), later switched on in CrewAI with memory=True (vector store for embeddings, SQLite for history).

PROCEDURE. (1) Define an LLM via crewai.llm.LLM with model, api_key (OpenRouter - one key for 300+ models instead of separate OpenAI / Anthropic / xAI / Google credits), base_url, temperature, max_tokens, timeout. (2) Define an Agent with role, goal, backstory, llm and tools. (3) Define a Task with description, expected_output, agent, optional context and output_file - in tasks.py. (4) Define a Crew with agents, tasks, verbose, process - in main.py. (5) crew.kickoff(inputs=...). Demonstrated first as a conflict detector on two sentences, then as content writer, code reviewer and customer-support agents. Install with uv (or pip): uv install crewai. CrewAI was chosen over the OpenAI Agents SDK, LangChain, LangGraph and DSPy for syntax simplicity and docs (which are themselves maintained by a CrewAI crew).

PROCEDURE. Log Analyzer uses CrewAI's built-in FileReadTool to extract errors from a .log file; Issue Investigator uses ExaSearch to find solutions; Solution Specialist has no tools and synthesises both into a step-by-step remediation plan. Downstream tasks receive upstream outputs through Task(context=[...]); a shared system_template string ('You are an expert DevOps engineer...') is passed to every agent like team-wide coding standards, layered under each role / goal / backstory. Each stage writes a markdown output file.

max_iter caps how many times an agent may re-invoke the model; max_execution_time (seconds) is a hard wall-clock cutoff that takes precedence - hence the worked example; max_rpm caps requests per minute to avoid provider throttling. Ishan calls these essential even for small systems: they are the difference between a bounded bill and a runaway loop.

PROCEDURE: import the decorator, apply @tool to a plain function (calculate_square, a currency converter, a CSV reader, a URL shortener), pass the function reference in Agent(tools=[...]). Real use: three yfinance-backed tools - get_current_stock_price, get_company_info, get_income_statements - so the financial agents fetch deterministic numbers instead of asking a model to remember them. Built-in tools (file / PDF / CSV / JSON readers, website scraping, research, database, DALL-E, code interpreter, AWS S3 reader / writer) come from crewai_tools and are listed on the docs' Tools Overview page.

Three reasons from the investment adviser. Twenty-plus fundamental parameters are not reliably obtainable by ad-hoc web search, so a dedicated Data Explorer with custom tools fetches them. Hallucination is unacceptable when money is involved, so numbers come from deterministic tools, not model recall (he cites the paper 'Why Language Models Hallucinate'). News and financials are independent, so they run in parallel and the whole pipeline is 35-40% faster. The Analyst agent, with no tools at all, is 'one of the most important' - pure interpretation of P/E, revenue, margins, risks.

PROCEDURE: wrap each independent single-agent crew (financial, news) in a function, submit both to concurrent.futures.ThreadPoolExecutor with executor.submit(), collect results, then run the sequential analysis crew (Analyst -> Financial Expert). Ishan avoids CrewAI's own async_execution flag, which 'breaks for weird reasons'. Measured live: parallel phase ~35 s, sequential phase ~68 s, ~100 s total.

Three tiers: 10-20 years - lead AI-first teams and governance; 3-9 years - ship agentic and RAG systems and automate legacy workflows; 0-3 years - become the organisation's generative-AI go-to person. Followed by recorded learner testimonials. Promotional framing rather than technique, recorded here as context for the program's positioning.