11  What to Learn to Build AI Agents (Roadmap)

A practical, ordered learning path — the same order used in Python ka Chilla: Build AI Agents by Codanics. Free playlist: https://www.youtube.com/watch?v=UE3o-7KRNXs&list=PLaQmdklCcX40

11.1 Stage 0 — Python you actually need (2–3 weeks)

Topic Why it matters for agents Used in this repo
Functions, modules, packages Every tool is a function; every agent is a module all tools.py, agents.py
Type hints & dataclasses / Pydantic Structured outputs, validation every BaseModel
pathlib, JSON, CSV Saving outputs, memory, audit trails main.py, memory.py
Virtual environments (venv), pip, requirements.txt Reproducible installs 02_agents_code/.venv
Environment variables & .env Secrets (API keys) never in code llm.py (python-dotenv)
Exceptions & retries APIs fail; agents must not crash tools.search_all
HTTP with requests Calling research databases tools.py
CLI with typer, pretty output with rich Usable tools without a UI every main.py
pandas / numpy basics Data agents Agent 3

11.2 Stage 1 — LLM fundamentals (1–2 weeks)

  1. How an LLM call works: system prompt, user message, tokens, max_tokens, temperature vs. effort, streaming.
  2. Prompting: role, task, constraints, examples, output format. Learn to write system prompts that define a persona and rules (see CITE_RULE in agents.py).
  3. Structured outputs: force JSON that matches a schema (Pydantic → JSON schema). This is the single most important skill for reliable agents.
  4. Context windows & cost: what to put in the prompt (only what is needed), prompt caching, choosing models (claude-opus-5 for quality, claude-sonnet-5 for bulk).
  5. Thinking/effort controls: adaptive thinking; low effort for screening, high for critique.

11.3 Stage 2 — Tools & agents (2–3 weeks)

  1. Tool use / function calling: give the model functions it can call; the tool-runner loop (call → execute → feed result back).
  2. The agent loop: Plan → Act (tool) → Observe → Repeat → Answer (ReAct).
  3. Pipelines vs. autonomous agents: this repo uses pipelines of small agents (predictable, testable). Fully autonomous loops come later.
  4. Multi-agent patterns: planner/worker, generator/critic, persona simulation (Agent 4), orchestrator (main.py).
  5. Memory: short-term (the prompt), long-term (files/DB), summaries (context_summary() in Agent 5).
  6. Retrieval (RAG): search → chunk → embed → retrieve → answer; citation grounding.

11.4 Stage 3 — Research-domain tools (1–2 weeks)

  • Scholarly APIs: arXiv, PubMed E-utilities, Semantic Scholar, Crossref, OpenAlex.
  • PDF text extraction (pypdf, GROBID), reference formats (BibTeX/RIS), DOIs.
  • Systematic-review methodology (PRISMA), evidence tables, PICO.
  • Basic statistics: t-tests, ANOVA, multiple-testing (FDR), effect sizes, cross-validation, leakage.

11.5 Stage 4 — Productising (3–4 weeks)

  1. Evaluation: golden sets, rubric grading, LLM-as-judge, regression tests for prompts.
  2. Web app: Streamlit (MVP) → FastAPI + React; background jobs; file uploads.
  3. Storage: SQLite/Postgres for papers, runs and memory; object storage for PDFs.
  4. Observability: logs of every prompt/response, token usage, cost per run.
  5. Safety & integrity: no citation without a source; AI-use disclosures; privacy (send summaries not raw data).
  6. Deployment: Docker, cloud VM/containers, scheduled runs, secrets management.
  7. Business: pricing, pilots, sales collateral (see 01_guides/07_go_to_market_and_pricing.md).

11.6 Stage 5 — Advanced (ongoing)

  • Autonomous tool-runner agents with approval gates; sandboxed code execution.
  • Managed / hosted agents with scheduled deployments; MCP servers for tools.
  • Fine-tuning vs. prompting; embeddings & vector DBs; graph-based citation analysis.
  • Multi-modal inputs (figures, PDFs, images) for extraction.

11.7 Suggested weekly plan (8 weeks)

Week Build
1 Python refresh; run agent_03 --demo --no-llm (no key needed)
2 Get an API key; run Agent 1 with --offline, then live
3 Modify tools.py: add OpenAlex; add a new field to Paper
4 Run Agent 2; edit evidence.csv; --from-evidence
5 Run Agent 3 on your own CSV; add a new test in analysis.py
6 Run Agent 4 on your notes; add a 4th reviewer persona
7 Run Agent 5 for 3 “weeks”; build a Streamlit status page
8 Write evals; pick one agent; make it a product (UI + pricing page)

Codanics — Python ka Chilla: Build AI Agents

  • Course: https://codanics.com/courses/python-ka-chilla-build-ai-agents/
  • Website: https://codanics.com · YouTube: @codanics
  • WhatsApp channel: https://whatsapp.com/channel/0029Va7nRDq3QxRzGqaQvS3r
  • Complete free course: https://www.youtube.com/watch?v=UE3o-7KRNXs&list=PLaQmdklCcX40