Skip to content
Enrollment is open — create your account and start with the free coursesCreate account
MASTERYCERTIFICATIONS

Independent professional certification

Mastery Certified: AI Agent Developer (Claude)

Shows that you can design, build and operate applications and agents on the Claude API: requests, models and errors; prompts, structured outputs and caching; tools and agent loops; MCP; evaluation, observability and safety. It is meant for software developers, data and ML engineers and technical leads who already write Python and work with HTTP/JSON APIs. You earn it by passing the graded exam and having your final project approved.

Take the examSee the preparation course
Exam
60 questions
Duration
120 minutes
Passing score
70 %
Final project
Required
Valid for
24 months
Preparation
30 hours

Exam syllabus

The domains the exam covers and their weight. Open each one to see its objectives.

  1. 1API fundamentals15 %
    • Structure Messages API requests (model, max_tokens, system, messages, content blocks) and explain why the API is stateless.
    • Interpret responses: content block types, usage fields and every stop reason, with the right handling for each.
    • Choose models by querying the Models API and by task needs; configure adaptive thinking and effort; avoid parameters current models reject.
    • Stream long outputs and large max_tokens requests and assemble the final message.
    • Handle errors by class: retry 429, 5xx and 529 with backoff and the retry-after header; never retry 400, 401 or 404; keep credentials out of code.
  2. 2Prompt and context engineering15 %
    • Write system prompts with a role, context and motivation, explicit actions, XML structure and examples.
    • Get machine-readable results with structured outputs (output_config.format) and strict tool use, within the schema limits, handling refusals and truncation.
    • Organize long context: documents first, the question last, quote grounding, metadata in tags; the Files API and citations.
    • Design for prompt caching: render order, breakpoints, TTLs, silent invalidators and verification through the usage fields.
    • Manage the context window: token counting, trimming that keeps tool pairs intact, compaction and context editing.
  3. 3Tool use and agent loops25 %
    • Define tools with precise names, descriptions that say when to call them, and JSON Schemas with enums, required fields, additionalProperties: false and strict mode.
    • Implement the agent loop: append the assistant content, run each tool_use, return every tool_result with its tool_use_id in one user message, stop on end_turn and cap iterations.
    • Handle parallel calls, tool failures (is_error), truncated tool calls and refusals inside a loop.
    • Use the SDK Tool Runner and decide when to own the loop.
    • Choose between a single call, a workflow (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer) and an agent.
    • Validate tool inputs before running them and make side-effecting tools safe to retry.
  4. 4MCP and integrations15 %
    • Explain MCP's architecture: hosts, clients and servers; JSON-RPC 2.0; tools, resources and prompts; sampling, roots and elicitation.
    • Read and build MCP messages: tools/list, tools/call, results with isError, protocol errors; the legacy initialize handshake versus per-request metadata.
    • Configure the MCP connector: mcp_servers plus mcp_toolset, the beta header, allowlists and denylists, and its limits.
    • Connect local MCP servers through client-side helpers or the Claude Agent SDK.
    • Choose between your own loop, the Tool Runner, Managed Agents and the Claude Agent SDK.
  5. 5Evaluation and observability15 %
    • Define success criteria and build eval sets from real traffic, edge cases and adversarial cases.
    • Choose and calibrate graders: code checks, rubric grading by a model, human review.
    • Run regression evals on every prompt, model or tool change and compare versions by category.
    • Log and trace requests (request id, model, stop reason, usage, latency, tool calls) with sensitive data redacted.
    • Measure cost and latency per completed task and reduce them with caching, Batches and effort.
  6. 6Safety, security and operations15 %
    • Recognize OWASP Top 10 for LLM Applications (2025) risks in an agent design.
    • Defend against indirect prompt injection: tool results are data, trusted instructions stay separate, outputs are validated.
    • Apply least privilege and human approval: read-only defaults, permission gates, limits, audit logs.
    • Handle data responsibly: minimization, retention, secrets, personal data in logs.
    • Operate safely: budgets, iteration caps, fallbacks for refusals and outages, incident response.

How to earn it

  1. Prepare

    Complete the preparation course: it covers every exam domain, with practice and a practice exam.

  2. Submit the final project

    Solve the course capstone case; the academic team reviews it with a rubric.

  3. Pass the graded exam

    Taken on the platform, with the stated time limit and passing score, by people enrolled in one of its courses.

  4. Receive your certification

    With a serial number, a QR code and public online verification anyone can check.

Independent certification by Mastery Certifications. Not affiliated with, endorsed by or an official certification of Anthropic. Claude is a product of Anthropic, PBC.