anthropics/agent-sdk-workshop

anthropics/agent-sdk-workshop

Language: Python

License: Apache-2.0

Stars: 36

Forks: 14

Open issues: 1

Created: 2026-02-27T18:14:02Z

Pushed: 2026-03-05T21:28:08Z

Default branch: main

Fork: no

Archived: no

README:

Agent SDK Workshop

Hands-on workshop materials for building production agents with the Claude Agent SDK.

No code writing required. You flip switches, pick from component lists, and write prompts. The SDK does the rest.

---

Quick start

# 1. Clone and enter
git clone agent-sdk-workshop
cd agent-sdk-workshop

# 2. Install (Python 3.10+)
pip install -r requirements.txt

# 3. Set your API key
cp .env.example .env
# → edit .env and paste your key after ANTHROPIC_API_KEY=

# 4. Verify everything works
./workshop check
# (on Windows: python workshop check)

# 5. Run the first exercise
./workshop demo

---

Structure

01-guided-demo/ One agent, four stages. Flip switches in config.py, re-run, watch it improve.
02-breakouts/ Pick a use case. Assemble an agent from pre-built tools and sub-agents.
extend/ Finished early? Recipes for adding your own tools, sub-agents, and breakouts.
docs/ Cheat sheet, FAQ, troubleshooting.

---

The ./workshop command

One entry point for everything:

| Command | Does | |---|---| | ./workshop check | Verify your setup before starting (Python version, API key, SDK) | | ./workshop demo | Run the guided demo agent | | ./workshop demo --show-prompt | See the full context the SDK is sending to the model | | ./workshop breakout | List available breakouts | | ./workshop breakout | Run a specific breakout | | ./workshop reset | Clear all memory files (fresh start) |

On Windows, prefix with python: python workshop check.

---

Part 1 — Guided demo

One agent. One task. Four runs. Between each run, flip one switch in config.py and watch the agent get better at the same job.

| Stage | Toggle | SDK primitive | What unlocks | |---|---|---|---| | 0 | *(all off)* | system_prompt | Just chat — no tools, no state | | 1 | ENABLE_TOOLS | @tool + mcp_servers | Agent can look things up | | 2 | ENABLE_SUBAGENTS | AgentDefinition + Task tool | Agent delegates to specialists | | 3 | ENABLE_MEMORY | hooks + a persistence tool | Agent remembers across restarts |

→ Open [01-guided-demo/GUIDE.md](./01-guided-demo/GUIDE.md) for the stage-by-stage walkthrough.

---

Part 2 — Breakouts

Pick a use case. Assemble an agent. You get a library of pre-built tools (6 categories, 19 tools) and 6 sub-agents. Each breakout's config.py is where you pick which to enable and write the system prompt.

| Breakout | Scenario | |---|---| | 00-warmup | Same briefing task from the demo, but assembled the breakout way | | chief-of-staff | Board meeting prep, email drafting, exec briefings | | customer-support | Ticket triage with KB-grounded responses | | sre-agent | Incident investigation — correlate deploys, metrics, runbooks | | account-intelligence | Pre-renewal account review — surface hidden churn risks | | freeform | Blank canvas — you have your own use case |

→ Open [02-breakouts/README.md](./02-breakouts/README.md) to pick one.

---

Requirements

• Python 3.10+ (the SDK uses modern type syntax)
• An Anthropic API key
• That's it. Every tool is backed by local mock data — no external services, no extra credentials.

---

If something breaks

1. Run ./workshop check — catches most setup issues 2. Check [docs/TROUBLESHOOTING.md](./docs/TROUBLESHOOTING.md) 3. Check [docs/FAQ.md](./docs/FAQ.md)

---

Going further

• [extend/](./extend/) — recipes for adding your own tools, sub-agents, and breakouts
• [docs/CHEATSHEET.md](./docs/CHEATSHEET.md) — every SDK pattern on one page
• Claude Agent SDK — the SDK itself, with more examples
• claude-cookbooks / claude_agent_sdk — deeper reference agents