The short version
Several agentic AI certifications exist, and they are mostly vendor-aligned. For building agents on Claude, the credential is Anthropic's Claude Certified Architect - Foundations (CCA-F), where Domain 1 (Agentic Architectures) is 27% of the exam - the single heaviest domain, and draws directly on Anthropic's own published agent-architecture guidance ↗. NVIDIA, Microsoft, and AWS run their own agentic-AI learning paths aligned to their platforms; choose the one matching the stack you build on. The CCA-F is 60 questions, 120 minutes, 720/1000 to pass, five weighted domains, a $125 fee, valid 12 months, launched 2026-03-12, with eligibility gated through the Anthropic Partner Network ↗ and proctored delivery via Pearson VUE.
Is there an agentic AI certification?
Yes - more than one. Unlike a single industry-wide standard, agentic AI certifications are largely tied to vendor platforms. That is not a flaw; it reflects how different the toolchains are. An agent on one platform uses different orchestration, tool, and deployment primitives than an agent on another, so the certifications specialise.
The practical consequence: there is no single "agentic AI certification" to chase. There is the right one for your stack. If you build on Claude, that is the CCA-F. If you build primarily on another vendor's platform, their aligned path will map more closely to your daily tools. The first decision is the platform, and the certification follows.
The agentic-AI certification landscape
Here is the landscape at a safe altitude, framed by vendor, focus, and what the credential proves. Vendor catalogs evolve, so treat specifics as directional (🟡 consensus).
| Vendor | Focus | What it proves |
|---|---|---|
| Anthropic (CCA-F) | Architecting agents on Claude | You can design agentic architectures, tools/MCP, operations, prompts, and reliability for Claude. D1 is 27%. |
| NVIDIA | Agentic-AI learning path on NVIDIA tooling | Competence within NVIDIA's agentic-AI stack and developer ecosystem. |
| Microsoft | Agentic-AI learning path on Microsoft platforms | Building and operating agents within the Microsoft AI ecosystem. |
| AWS | Agentic-AI learning path on AWS services | Building and deploying agents using AWS AI services. |
The honest read of the table: each credential proves competence on its own platform. None is universally "the agentic AI certification." If your work runs on Claude, the CCA-F (architect) or CCD-F (developer) is the one whose weighting and content match your daily reality, and the rest are adjacent learning rather than direct substitutes.
What the CCA-F tests on agentic architecture
Domain 1 (Agentic Architectures, 27%) is the heaviest domain on the exam, and it breaks into four clusters. The exam tests each through scenario stems - a task plus a constraint, and you pick the architecture.
- Multi-agent patterns. When one agent is enough and when to split work across several. Stems give you a workload and ask whether a single agent, a pipeline, or a coordinator-and-workers shape fits best.
- Orchestration. How a coordinator delegates subtasks, aggregates results, and handles partial failure. The exam probes whether you can keep an orchestrator's context manageable as the agent count grows.
- Agent lifecycle. Plan, act, observe, recover, terminate. You are expected to reason about where a loop should stop, how to recover from a bad observation, and how to avoid runaway loops.
- Subagents. Spawning focused workers with their own scoped context, passing them the right inputs, and reining in context bloat. This cluster overlaps with context management in D5.
As with the rest of the exam, the test is judgment over recall. You choose the architecture that satisfies the constraint - latency, cost, reliability, or context budget - rather than naming a pattern. Tools are the actuators of any of these designs, which is why D1 connects straight into tool calling and MCP.
How to prep
D1 carries the most weight, so it deserves the most study time. The fastest route through it is to build, then drill.
- Build a multi-agent system. Stand up an orchestrator with two or three subagents on a task that genuinely requires delegation. Watch the full lifecycle and where it breaks.
- Practise the architecture choice. For each scenario, decide single-agent versus multi-agent and justify it against the constraint. Run the agentic tool design scenario to connect architecture to tools.
- Drill with timed questions. Use the practice questions to rehearse scenario stems until the architecture choice is reflexive. Start with the agentic loops concept. The CCAR-F certification page shows the complete five-domain blueprint.
- Cover the adjacent domains. Architecture decisions cascade into tools (D2), operations (D3), and reliability (D5). Study them as extensions of D1, not as separate silos.
How this maps to the CCA-F exam
References on this site
- Exam guide - CCA-F format, the five domains and their weightings, scoring, and validity.
- Registration and cost - how to register through the Anthropic Partner Network and Skilljar.
- Agentic tool design scenario - applied practice connecting architecture to tools.
- Tool calling concept page - the loop that turns architecture into action.
- MCP concept page - the protocol behind the tools an agentic architecture relies on.
Frequently asked questions about agentic AI certification
Is there an agentic AI certification?
Yes - several, and they are mostly vendor-aligned. The right one depends on the platform you build agents on. NVIDIA, Microsoft, and AWS each run agentic-AI learning paths or certifications tied to their own ecosystems. For building agents on Claude and Anthropic, the primary credential is the Claude Certified Architect - Foundations (CCA-F), where Domain 1 (Agentic Architectures) is the single heaviest domain at 27%; the Developer cert (CCD-F) also covers agent-building (Agents & Workflows). Anthropic runs four Claude certifications in total.
Which agentic AI certification should I take?
Match the certification to the stack you actually build on. If you ship agents on Claude, two certs fit by role: the CCA-F (architect-level - agentic architecture, tool and MCP design, Claude Code configuration, reliability) and the CCD-F (developer-level - the Agent SDK and API, building and shipping agents, with its own Agents & Workflows domain). Pick CCA-F if you design the system, CCD-F if you build it. If you build primarily on another vendor's platform, their aligned certification will map more closely to that toolchain. The wrong move is collecting certifications that do not match your daily stack.
What is the agentic-AI certification landscape?
At a safe altitude: Anthropic's CCA-F certifies architecting agents on Claude, with Domain 1 (Agentic Architectures, 27%) as the heaviest domain. NVIDIA, Microsoft, and AWS offer agentic-AI learning paths and certifications aligned to their own platforms and developer tooling (🟡 consensus - vendor catalogs evolve). Each proves competence on its own stack. They are complements, not substitutes, and your choice should follow the platform you deploy on.
Why is the CCA-F the right credential for Claude agents?
Because it is Anthropic's own certification for designing agentic systems on Claude, and its weighting reflects that focus. Domain 1 (Agentic Architectures) alone is 27% of the exam, the largest of the five domains, and Domains 2 through 5 cover the tool, operations, prompting, and reliability skills that an agent on Claude actually requires. No other vendor's certification is calibrated to the Claude stack the way the CCA-F is.
Is a prompt engineering certification the same as the Claude certification?
No, and they are not substitutes. A standalone prompt engineering certification is narrow: it tests prompting in isolation. On the CCA-F, prompt engineering is Domain 4 (20% of the exam), and it sits alongside agentic architecture, tool and MCP design, Claude Code configuration, and reliability. If you want to prove you can build and operate Claude systems, not just write prompts, the CCA-F is the broader, architecture-level credential.
What about an LLM certification or AI engineer certification?
Those are usually vendor-neutral and broad, covering general model concepts or a wide engineering toolchain. The CCA-F is the opposite: Claude-specific and architecture-focused. If your daily work is building agents and tools on Claude, the CCA-F maps directly to it; a general LLM or AI-engineer certification proves breadth but not depth on the Claude stack. Pick the one that matches what you actually ship.
What does the CCA-F test on agentic architecture?
Domain 1 (Agentic Architectures, 27%) covers multi-agent patterns (when to use one agent versus several), orchestration (how a coordinator delegates and aggregates), the agent lifecycle (plan, act, observe, recover, terminate), and subagents (spawning focused workers and managing their context). Stems are scenario-driven: you are given a task and a constraint, and you choose the architecture that fits rather than reciting definitions.
How many CCA-F questions are about agentic architecture?
Domain 1 is 27% of the 60-question exam, so roughly 16 questions touch agentic architecture - the largest share of any domain. If you study only one domain deeply, study this one. It also reinforces the others, because orchestration decisions cascade into tool design (D2), operations (D3), and reliability (D5).
How is the CCA-F structured?
Five weighted domains: D1 Agentic Architectures 27%, D2 Tool Design + MCP Integration 18%, D3 Claude Code Configuration & Workflows 20%, D4 Prompt Engineering 20%, and D5 Context + Reliability 15%. The exam is 60 questions in 120 minutes, with a 720/1000 passing score. It launched 2026-03-12; eligibility is gated through the Anthropic Partner Network and the proctored exam is delivered via Pearson VUE. It costs $125 ($0 for eligible partner employees) and is valid for 12 months.
How do I prep for the agentic-architecture portion?
Build before you read. Stand up a small orchestrator with two or three subagents, give it a task that requires delegation, and watch the lifecycle: planning, acting, observing, recovering, terminating. Then drill scenario questions that ask you to pick single-agent versus multi-agent, or the right orchestration pattern under a constraint. Architecture is judgment, and judgment comes from having shipped something.
Do other vendors' agentic certifications transfer to the CCA-F?
Partially. The underlying concepts - planning loops, tool use, orchestration, evaluation - are portable, so a vendor certification builds genuine intuition. But the CCA-F tests Claude-specific patterns, MCP integration, and Anthropic's framing of context and reliability, so you still need to study the CCA-F material directly. Treat prior agentic experience as a head start, not a shortcut past preparation.
Is the CCA-F worth it if I already build agents daily?
Often yes, because the credential signals validated competence to employers and clients who cannot inspect your code. Practitioners usually find Domain 1 (Agentic Architectures, 27%) the most natural and the operations and reliability domains the most clarifying, since day-to-day building can leave blind spots in those areas. The exam is calibrated for working architects, not beginners.
How does this map to the CCA-F exam?
Primarily Domain 1 (Agentic Architectures, 27%), the heaviest domain, for multi-agent patterns, orchestration, lifecycle, and subagents. It connects directly to D2 (tools as actuators), D3 (operating those agents), and D5 (keeping them reliable). Study D1 as the spine of the exam and the other domains as the limbs that hang off it.