A hands-on CrewAI lab: progressive modules introduce Agents, Flows, Conversational Flows, MCP, tools, human review, and deployment. The use cases are intentionally generic — triage, research, permissioned assistants — so the lab can be reused with any customer. All data is synthetic; nothing connects to production systems or sends email.
- Python 3.10–3.13
uv— installs and runs everything- A model-provider API key (e.g.
OPENAI_API_KEY) matching theLAB_MODELyou set
The repository pins CrewAI 1.15.5 so every attendee is on the same API surface — Conversational Flows are experimental, so keep the pin.
cp .env.example .env # add your model-provider key
uv sync --extra dev # installs the lab
uv run python modules/02-triagex-flow/solution_extended.py --standardThat last command is the smoke test: a full deterministic flow run with no API keys or network — if it prints a triage summary, the lab is ready.
| Module | Run it | Extra credentials |
|---|---|---|
| 01 · Agent Basics | uv run python modules/01-company-research-agent/solution.py [company] |
Model key. Firecrawl's keyless tier works; add FIRECRAWL_API_KEY for higher limits |
| 02 · Flow Foundations (TriageX) | uv run python modules/02-triagex-flow/basic.py — deterministic, no keys |
— |
| 02 · extended (router + HITL) | uv run python modules/02-triagex-flow/solution_extended.py --standard (no keys) or without the flag (model key, interactive review) |
Model key for the incident path |
| 02 · deployable scaffold | cd modules/02-triagex-flow/triagex_flow && crewai install && crewai run |
Model key + CREWAI_PLATFORM_INTEGRATION_TOKEN (GitHub connected in CrewAI AMP). Linear writes are opt-in: LINEAR_API_KEY, LINEAR_TEAM, SYNC_TO_LINEAR=true |
| 03 · Conversational Routing | cd modules/03-conversational-routing/permissioned_chat && crewai install && crewai run — switch identity via CHAT_PROFILE (analyst/viewer/guest) |
Model key. Firecrawl keyless tier is fine |
| 04 · Self-Improving Triage (human_feedback + memory) | cd modules/04-triagex-flow-imp/triagex_flow_imp && crewai install && crewai run — pauses for your verdict; run twice to see (N learned) |
Same as the module 02 scaffold; verdict memories live in ./storage |
Earlier modules covering tools+MCP, flow memory, and HITL outreach deployment
are parked in archive/ and can be revived per customer.
Every module directory has a README.md (framing, exercise, checkpoint) and a
basic.py with intentional TODOs. Modules 02 and 03 also ship slide decks —
open modules/02-triagex-flow/deck.html and
modules/03-conversational-routing/index.html in a browser.
Modules 02 and 03 carry their complete solution as a standard
crewai create flow project — self-contained src/<name>/main.py, its own
pyproject.toml with kickoff/plot/run_with_trigger entry points — so
each deploys to CrewAI AMP as-is:
modules/02-triagex-flow/triagex_flow/— triage GitHub issues/PRs from a date range → Linearmodules/03-conversational-routing/permissioned_chat/— permission-aware conversational assistantmodules/04-triagex-flow-imp/triagex_flow_imp/— module 02 + human verification (@human_feedback) and learned urgency (remember/recall)
Scaffolds are standalone projects: crewai install inside one creates its own
venv (fast — uv hardlinks from cache). They are intentionally not part of the
root environment. Each scaffold has its own .env.
External writes never happen by default anywhere in the lab: Linear publishing requires explicit opt-in, and GitHub access is read-only.
uv run python modules/02-triagex-flow/solution_extended.py --standard # offline flow run
uv run ruff check .
uv run python -m compileall -x '\.venv' src modulesThe former case-management and renewal modules are parked in
archive/. More on CrewAI skills:
https://www.skills.sh/crewaiinc/skills