Agent SDK Workshop

anthropic‘s hands-on teaching repository for the claude-agent-sdk (Apache 2.0, Python 3.10+, ~56★ / 17 forks, one commit on main as of 2026-08-10). It is the first source in this spoke that shows the vendor teaching the harness-as-library rather than documenting or selling it, and what it teaches is an ordering.

The premise is stated flatly: “No code writing required. You flip switches, pick from component lists, and write prompts. The SDK does the rest.”

The four stages

Part 1 runs one agent at one task through four stages, each a boolean in config.py:

Part 2 is six scenario breakouts — warmup, chief-of-staff, customer-support, sre-agent, account-intelligence, freeform — over 19 pre-built tools in 6 categories plus sub-agents, all running against mock local data so nothing external is needed. extend/ holds recipes for custom tools and agents; docs/ a cheatsheet, FAQ and troubleshooting. Setup is a clone, pip install, an ANTHROPIC_API_KEY in .env, then ./workshop check and ./workshop demo.

Why the ordering matters

This spoke assembled its layer list bottom-up, from artifacts that each argued for one layer at a time. The workshop is the same stack stated by the vendor, in the order it wants you to add it: prompt, then tools, then delegation, then memory. Capability first, coordination second, persistence last.

That is a claim about dependency, and it is legible because the switches are independent. Nothing in the SDK stops you from enabling memory before tools. The workshop puts memory last anyway, which reads as a judgment that persistence is worth least until there is something worth persisting — consistent with what iai-pme and jcode found the hard way, and stated here as a curriculum decision rather than as a finding.

The pedagogy is also the argument. Four boolean flags say agent-building is configuration, not programming, which is the harness-as-library thesis pushed to its end: the SDK owns the loop, the builder supplies switches and prose. Whether that survives contact with a real deployment is not something a workshop can answer.

The scenario list is a market read

The six breakouts are Anthropic’s picture of where SDK agents go: an executive assistant, a support desk, an SRE responder, sales account research. Four of six are internal-operations roles rather than coding, which is a quiet departure for a spoke whose corpus is dominated by coding harnesses. The sre-agent scenario is the seam with ../platform-ops-wiki (which owns AIOps and incident response as practice); here it is a teaching fixture, not an operational account.

What it does not do

No evaluation, and the structure implies one. Each stage is presented as an improvement on the last, and the repository measures nothing — a natural consequence of a workshop, but worth naming because the staging asserts monotonic gain and this corpus keeps failing to find it. cc-thinking-skills, context-files-ablation-paper and jetbrains-ponytail-tested all measured a structural layer and found process effects without correctness effects. The workshop’s stage 1 (tools) is the one addition none of those studies covered and the one most likely to be real: a tool supplies facts the model does not have, where a context file supplies facts it mostly did.

Mock data throughout. The tools return canned local fixtures. That is the right call for a classroom — no keys, no rate limits, no flaky third parties — and it means nothing here is evidence about behaviour against real systems.

One commit, 56 stars. New enough that its stability is unknown, and thin enough that adoption says nothing yet.

Tier

T1 as an Anthropic first-party artifact, per ../QUALITY.md. Tier and trust are separate axes: this is authoritative about what the SDK offers and about how its vendor thinks it should be learned, and it is evidence of nothing about whether agents built this way work.

Related

claude-agent-sdk · claude-code · agent-orchestration · agent-memory · durable-agents · model-context-protocol · agentic-coding-harness · anthropic · claude-managed-agents