Anthropic 'Code with Claude': Workshop Collection for Agent & Eval Practicals

This repository collects Anthropic's 'Code with Claude' hands-on workshops covering model selection, multi-agent design, memory and evaluation—useful for learning agent engineering and eval-driven development—but it is declared unmaintained and heavily relies on Anthropic services, posing migration and long-term use risks.

GitHub anthropics/cwc-workshops Updated 2026-07-18 Branch main Stars 1.6K Forks 489

Mixed/Unknown Claude Managed Agents Multi-Agent Systems Eval-Driven Development Teaching/Workshop Materials Apache-2.0 Read-only / No contributions

💡 Deep Analysis

How to use the repo's model-audit and fast-eval workflows to pick models and inference parameters?

Core Analysis ¶

Project Positioning: The repo’s rightmodel/ and agent-battle/ patterns advocate using short-cycle fast-eval probes to filter configurations, followed by longer benchmarks on shortlisted candidates to find the best quality/cost/latency trade-off.

Technical Features ¶

Short-cycle decision probes: Fast loops (~30s) enable frequent configuration iterations.
Systematic parameter sweeps: Sweep models and inference params (temperature, extended thinking, effort).
Dual evaluation signals: Programmatic metrics (structure, latency, cost) plus LLM-as-judge for subjective quality.

Practical Steps (How-to)¶

Create a small eval suite (10–20 representative tasks) for fast-eval.
Run parallel sweeps: Explore cost-sensitive parameters and record latency/token usage.
Select top-N configs: Validate with larger (100+ task) benchmarks and LLM judges for stability.
Integrate into CI: Automate metrics to catch regressions.

Important Notice: LLM-as-judge can be gamed by prompts; keep programmatic metrics and update rubrics regularly.

Summary: Fast-eval plus systematic sweeps is a cost-effective selection strategy suitable for early iteration and final production validation.

85.0%

For engineers new to the repo, what is the learning curve, common pitfalls, and how to get up to speed quickly?

Core Analysis ¶

Project Positioning: Targeted at engineers with ML/software background, the repo expects familiarity with web stacks and infra. Engineers unfamiliar with Anthropic constructs face a steeper learning curve.

Common Pitfalls ¶

Platform dependencies & permissions: Full reproduction requires access to Claude Managed Agents APIs and API keys.
Operational complexity: Multiple local services (MCP, memory service, local tool endpoints) are easy to misconfigure.
Evaluation misinterpretation: Mixing programmatic metrics and LLM judges without clear rubrics can mislead tuning.

Quick Start Recommendations ¶

Run smallest workshop first: Start with ship-your-first-managed-agent and implement the seven small functions.
Validate core dependencies: Ensure API keys and minimal local endpoints are working before adding complexity.
Use fast-eval to shorten feedback loops: Frequently validate changes to avoid long benchmark cycles.
Keep rubrics & programmatic metrics explicit: Maintain reproducible evaluation standards when tuning.

Important Notice: The repo is marked “not maintained”; expect to patch SDK or dependency changes yourself.

Summary: A staged, small-to-large approach with fast-eval reduces ramp-up time, but platform dependency and maintenance risks remain.

85.0%

How can the repo's machine-readable DOM contracts and CI-style runtime verification be implemented in real engineering practice?

Core Analysis ¶

Project Positioning: The repo recommends contractizing front-end components and agent interactions (machine-readable DOM contracts) and embedding these contracts into CI and runtime validation to ensure consistent agent–UI behavior end-to-end.

Implementation Highlights ¶

Contract expression: Define expected fields, events, and state transitions with JSON schema or similar formats (e.g., button semantics, streaming events, tool-call structures).
CI-level validation: Run contract tests in end-to-end suites by simulating agent outputs and asserting DOM contract conformance.
Runtime validation: Validate agent-generated events/DOM snippets against schemas in production; on violations, trigger alerts, intercept, or degrade to safe flows.

Practical Recommendations ¶

Contractize critical interactions first: Start with high-risk paths (tool calls, exports, payment flows).
Automate contract tests: Add contract tests to CI to prevent regressions from frontend or agent updates.
Design failover & mitigation: When contract checks fail, prefer safe degradation or human-in-the-loop review to avoid bad actions.

Important Notice: Contracts must be maintained with product evolution—use versioning to handle breaking changes.

Summary: Machine-readable DOM contracts plus CI/runtime validation increases agent–UI reliability but requires investment in schema management, test coverage, and change governance.

85.0%

✨ Highlights

Provides multi-scenario, modular hands-on workshop materials covering end-to-end examples
Covers model selection, agent decomposition, memory, evaluation and production-ready examples
Repository is declared unmaintained and not accepting contributions; community engagement is limited
Examples depend heavily on Anthropic private services/APIs; migration and long-term operation carry risk

🔧 Engineering

Engineering-focused multi-agent and skill-based practical examples to learn architecture and layered design
Includes end-to-end workshop materials and runnable examples from design through evaluation

⚠️ Risks

No active contributors, no formal releases and sparse commit records; maintenance and security updates are not assured
Content depends on closed/hosted Claude services and internal APIs; reproducing or running outside that ecosystem may be difficult

👥 For who?

Suitable for AI engineers and researchers who need hands-on examples to understand agent engineering and evaluation workflows
Can serve as reference material for internal training, teaching, and proofs-of-concept, but not as a production dependency