Home › Certifications › CCAR-F for Generative AI Developers
Claude Certification Guide · 2026CCAR-F for Generative AI Developers: From RAG App to Agent
The Claude architecture credential read as a step up, not a sideways move. Why Prompt Engineering & Structured Output is the bridge domain you are already halfway across, what changes when the model's answer becomes another component's input, why one unreliable JSON field breaks a whole chain, and how it pairs with AI-103.
Authorized Training & Technology Partners
Claude
Claude Certified Architect – Foundations (CCAR-F)
The credential this guide covers
Microsoft
AI-103 — Azure AI Apps & Agents Developer Associate
The other credential in this programme — read the AI-103 guideIf you have shipped a RAG application, you are closer to CCAR-F than you think in one domain and further than you think in another. Prompt Engineering & Structured Output (20%) is work you already do. But an agent turns the model's answer into another component's input — and once output is consumed by code rather than read by a person, “usually correct” stops being good enough.
This is the generative-developer reading of CCAR-F. For the credential explained generally see our CCAR-F certification guide; for exam tactics, how to pass CCAR-F. What follows is about the specific gap between a generative application and an agentic one, and how the exam tests it.
What actually changes between a RAG app and an agent
Both call a model. That is where the similarity ends, and misreading the difference is why capable generative developers underestimate this exam.
| Generative application | Agentic system | |
|---|---|---|
| Shape | Retrieve, then generate once | Decide, act, observe, decide again — until done |
| Who reads the output | A person | Your code, then the model again |
| A malformed response | Looks slightly odd to a user | Breaks the next step, and everything after it |
| Failure | One bad answer | A loop that will not terminate, or acts on a wrong premise repeatedly |
| Cost | Predictable per request | Depends how many steps the model chooses to take |
| Testing | Compare against expected answers | Evaluate a trajectory — often several valid ones exist |
In a generative app, a prompt that works 95% of the time is a good prompt — the user shrugs at the other 5%. In an agent, 95% per step compounds. Across eight steps that is roughly a 66% chance of a clean run. Structured output with validation is not polish here; it is what makes multi-step work viable at all, and it is why the domain carries 20%.
The bridge domain: prompting and structured output
This is where your existing skill converts fastest into exam marks — provided you upgrade it from “getting good answers” to “getting parseable, reliable answers every time”.
Schema-constrained, not schema-requested
Asking politely for JSON in the prompt is not the same as constraining the output to a schema. The first fails occasionally and unpredictably; the second is enforceable. Know which mechanism you are relying on — scenario questions turn on exactly that distinction.
Structured outputValidate, then decide what failure means
Parsing is not validation. A response can be valid JSON with a nonsensical enum value or an out-of-range number. Decide up front: retry, repair, escalate, or fail loudly? An agent that silently proceeds on bad data is worse than one that stops.
ReliabilityPrompts are production configuration
Version them, review them, test them. In an agent, a prompt regression is a production incident affecting every request that touches it — there is no gradual rollout by default. Generative developers who keep prompts in a notebook get caught by this.
Change controlSystem prompt design carries the architecture
In a single-turn app the system prompt sets tone and scope. In an agent it encodes the operating rules — when to use which tool, when to stop, when to ask a human. Much of what looks like orchestration logic actually lives here, which is why the domains overlap.
Prompt engineeringYour RAG work is not wasted
Retrieval becomes a tool the agent calls rather than a fixed pipeline stage — and everything you learned about chunking, hybrid search and reranking still applies. What is new is that the agent decides whether and when to retrieve, which is a design question you did not previously have to answer.
What carries overThe 27% you have not met
Agentic Architecture & Orchestration is the largest domain and the genuinely unfamiliar one: topologies, delegation, bounded loops, partial failure. Do not let the comfortable 20% absorb your study time. Budget by blueprint, not by confidence.
Where to spend the hoursCCAR-F domains and weightings
CCAR-F exam blueprint — share of content by domain
Official Anthropic weightings, Claude Certified Architect – Foundations
Source: Anthropic Partner Academy, CCAR-F exam blueprint. Percentages total exactly 100%. A generative developer typically arrives strong on the 20% prompting domain and with some of the 15% context domain, and starts near zero on the 27% orchestration domain.
CCAR-F at a glance
| Attribute | Detail |
|---|---|
| Credential | Claude Certified Architect – Foundations (CCAR-F) |
| Issued by | Anthropic |
| Format | 60 scenario-based items; 4 scenarios drawn from a published bank of 6 |
| Duration | 120 minutes, proctored |
| Scoring | Scaled 100–1000; 720 to pass |
| Exam fee | US$125 |
| Delivery | Pearson VUE — online proctored or at a test centre |
| Validity | 12 months from the date awarded |
| Paired with | AI-103 (Azure AI Apps and Agents Developer Associate) in the 360DT Generative AI Developer programme |
| Note on naming | Early material used CCA-F; the current Pearson VUE exam code is CCAR-F. CCFA is not an official code |
Who this angle is for
- Developers who have shipped a RAG or assistant feature and are now being asked to make it take actions rather than only answer.
- Prompt-heavy engineers who want the discipline that turns prompting from craft into something reviewable and testable.
- Anyone whose generative feature keeps breaking downstream because the model's output is usually the right shape.
- Anyone pairing it with AI-103 — the Azure platform half and the vendor-neutral architecture half of the same job.
Where it will not carry you
- Prompting skill alone will not pass it. That is 20%; orchestration is 27% and Claude Code another 20%.
- It is not about retrieval quality. Chunking, embeddings and reranking are AI-103 territory — here retrieval is simply one tool among several.
- It expects you to have built an agent, not only read about one. Scenario questions are unforgiving of theory-only preparation.
What the step from generative to agentic is worth
Building a RAG demo is now a common skill. Building an agent that takes actions and does not misbehave is not — and that is the gap the upper bands are paying for.
The AI developer progression
Entry
Mid
Senior
Lead / Architect
Generative and agentic developer pay by market
India and the US shown as two separate charts, because rupee and dollar bands are different measures on different scales and should never share an axis.
India — annual CTC
AI and generative AI developer roles, ₹ lakh per annum
Bar length maps the upper bound of each band against a ₹53 L scale. Bands compiled from published Indian AI engineer ranges; product companies and global capability centres sit at the top of each band.
United States — annual base
AI developer and agentic specialist roles, US$ thousands
Bar length maps each figure against a $200K scale. Bands compiled from published US AI engineer and prompt/agentic specialist ranges; base pay only, excluding bonus and equity. The top band is wide because the specialism is young and titles are inconsistent.
These are role bands, not certification outcomes. No published figure attaches a salary increase to holding CCAR-F. The 66% figure is plain arithmetic — 0.95 to the power of 8 — illustrating why per-step reliability matters in a multi-step system. It describes compounding error, not pay.
Compensation figures are compiled from independent, publicly available industry sources and are shown for role context. They are not a guarantee of pay in any specific market, company or outcome, and 360DT does not promise a salary result from any certification or programme.
Your 8-week roadmap, from a generative start
Light on prompting because you have most of it, heavy on orchestration because you have none of it. Assume 6–8 hours per week and build throughout.
Read all six scenarios, then upgrade your output discipline
Anthropic publishes the scenario bank and draws four of six, so read them all first. Then take a prompt you already use and make its output schema-constrained and validated. Decide explicitly what happens on a validation failure.
Prompts as production configuration
System prompt design, few-shot construction, decomposition. Put prompts under version control and write tests. Completes the 20% domain — the one you are closest to, and the fastest marks available to you.
Your first loop
Build plan-act-observe by hand, without a framework. Add a step limit, a timeout and a stop condition. Then break it on purpose: make it loop forever, then stop it properly. This is the week the difference from a RAG app becomes concrete.
Tools, including retrieval as a tool
Boundaries, descriptions, parameter design, readable errors. Convert your existing RAG pipeline into a tool the agent may call — and notice that the agent now decides whether to retrieve. That shift is the exam's view of your old work.
MCP
Servers, transports, capabilities. Stand one up and connect it to two clients. Completes the 18% domain, and gives you an integration pattern that outlives any single model choice.
Multi-agent topologies
Supervisor, pipeline, peer delegation. Hand-off and shared state. And the judgement worth the most marks: when one agent is simply better than three. Coming from single-turn apps, your bias toward simplicity is an asset here.
Claude Code and context reliability
Configure Claude Code into a real workflow and use it daily — 20% of the exam assumes you have. Then context: window strategy, summarisation, memory, and spotting long-run degradation before a user does. Completes the last 35%.
Scenario drills and timed mocks
Re-read all six scenarios and write the architecture and trade-off you would defend for each. Then full 60-question, 120-minute mocks scored by domain. Any domain under 70% gets a re-study day. Then book.
Read the final sentence of each scenario stem first — it usually names the constraint being optimised for, and that decides between two workable designs. Coming from generative work, watch one specific reflex: the answer is rarely “write a better prompt”. Where a structural fix — a bounded loop, a validated schema, a simpler topology — removes the failure, that is what is being looked for.
Learn from Anthropic Authorized Instructors
Certification prep at 360DT is delivered live by AWS, Microsoft, Cisco and NVIDIA certified trainers and Anthropic Authorized Instructors — our full international faculty.
Bipeen
AWS, Microsoft, and Anthropic Certified Trainer
AI, machine learning, and cloud transformation expert with more than 25 years of experience. A renowned global corporate trainer and conference speaker known for practical, outcome driven sessions.
Florian Garcia Compte
Cisco Certified Systems Instructor (CCSI)
Partner Director at AVAANZA FORMACION, a Cisco Learning Partner based in Madrid. Certified Cisco Systems Instructor (CCSI No. 21053) specializing in Cisco Data Center (ACI/SDN), Nexus, and Wireless, delivering official Cisco certification courses including CCNA, CCNP Enterprise, and CCNP Data Center.
Shantanu Pandey
Microsoft, Google, NVIDIA and HPE Certified Trainer
Director of Engineering at Meteoros Automation, ranked in Microsoft's Top 100 Trainers of 2025, holding MCT, Google Cloud, NVIDIA and PeopleCert credentials, with 500+ trainings delivered since 2011.
Vikas Mittal
Google and Anthropic Authorized AI Instructor
Veteran technology trainer with more than 25 years of rich industry experience spanning Google, Microsoft, and emerging AI platforms. An exceptional mentor recognized for proven teaching excellence.
Deep
Anthropic Authorized Instructor and AI Cloud Specialist
Seasoned technology leader with over 20 years of experience across AI, cloud computing, and enterprise training. Known for simplifying complex ideas and delivering practical, real world learning.
Sid
Anthropic Authorized Instructor and Azure Data Engineer
Dynamic corporate trainer and consultant with more than 10 years of experience across Microsoft Fabric, Azure Data Engineering, and modern cloud platforms. Focused on clear, practical learning that sticks.
Arshad Ahmad
Microsoft Certified Trainer (MCT), Cybersecurity & Power Platform Expert
Technology Trainer with 15 years in strategic client delivery, holding MCT, Azure Solutions Architect Expert, Cybersecurity Architect Expert and Power BI Data Analyst credentials, with 200+ trainings delivered.
Sravia
NVIDIA Authorized Instructor and AI Data Specialist
Passionate AI and data evangelist with over 10 years of experience across machine learning, MLOps, and cloud native technologies. Dedicated to delivering impactful, engaging enterprise training programs.
Johan
Microsoft Business Applications MVP and Microsoft Certified Trainer
D365 Customer Service Lead – Europe at Avanade, and founder of the Power Platform School. A Microsoft Business Applications MVP and Microsoft Certified Trainer based in London, specializing in Dynamics 365 Customer Service and Power Platform.
Akim Nyamande
Microsoft, CompTIA, Juniper and Cisco Certified Trainer
IT training facilitator with over five years of experience and certifications across Microsoft, CompTIA, Juniper, and Cisco. Previously a Network Administrator before moving into technical training, now delivering hands-on courses in networking, systems administration, and cybersecurity fundamentals.
Ali El Khatib
Cisco Certified Systems Instructor (CCSI) and Microsoft Certified Trainer
Infrastructure Engineer at RHUH with 13 years in the training field, holding CCSI, MCSE, CCNP Routing and Switching, CCNP Security, CompTIA and Microsoft Azure certifications.
Generative AI Developer Program
The 360DT Generative AI Developer Program runs the full AI-103 blueprint — Foundry, grounding and RAG on Azure AI Search, agents and orchestration, safety and evaluation, multimodal and information extraction — then carries on into Claude Certified Architect – Foundations (CCAR-F). You finish with five shipped systems, including a grounded enterprise assistant, a chunking and retrieval benchmark, and an evaluation and safety harness.
Explore the Generative AI Developer Program
CCAR-F for generative AI developers — frequently asked questions
I have shipped RAG applications. How big is the jump to agents?
Bigger than it looks in one respect. A RAG app generates an answer a person reads; an agent generates output your code consumes and acts on. Once that is true, output reliability stops being polish and becomes structural — which is why Prompt Engineering & Structured Output is 20% of the exam. Your retrieval work still counts; it just becomes a tool the agent may call.
Why is a 95%-reliable prompt not good enough?
Because it compounds. In a single-turn app, 95% means one user in twenty sees something odd. Across an eight-step agent run, 0.958 is about 66% — roughly a one-in-three chance of a broken run. Schema-constrained output with validation is what makes multi-step work viable.
Is asking for JSON in the prompt the same as structured output?
No, and the distinction is examined. Requesting JSON in the prompt is a hope; constraining the response to a schema is enforceable. And parsing is not validation — a response can be perfectly valid JSON with a nonsensical enum or an out-of-range number. Decide explicitly whether you retry, repair, escalate or fail.
Which domain will I find hardest?
Agentic Architecture & Orchestration, at 27% — the largest domain and the one with no equivalent in single-turn generative work. Topologies, delegation, bounded loops, partial failure and recovery. Budget your hours by the blueprint rather than by what feels comfortable.
Does my RAG and retrieval work still count?
Yes. Chunking, hybrid search and reranking all still apply — retrieval simply becomes one tool among several rather than a fixed pipeline stage. What is new is that the agent decides whether and when to retrieve, which is a design decision you did not previously have to make.
How is this different from your other CCAR-F guides?
Same exam, different reader. This one is for developers moving from generative applications to agentic ones. Our CCAR-F certification guide explains the credential, how to pass CCAR-F is exam strategy, the cloud developer guide takes an integration lens, and the MLOps guide an operations one.
Do I need AI-103 as well?
Not as a prerequisite. They cover different halves: AI-103 is the Azure platform side — Foundry, grounding, multimodal, extraction, safety — and CCAR-F is the vendor-neutral agent architecture side. Our Generative AI Developer programme prepares for both in eight weeks.
Are the exams included in the programme fee?
No. Certification exams are administered independently by Microsoft and Anthropic and are not included in programme tuition. The fee of ₹24,999 covers the live training, the five shipped systems and certification preparation; you book and pay for each exam directly with its issuer.
Sources and further reading
- Anthropic Partner Academy — Claude Certified Architect – Foundations exam blueprint (five domains with exact weightings, 60 scenario-based items, 120 minutes, scaled 100–1000 with a 720 pass mark, six-scenario bank with four presented per sitting, US$125 fee, 12-month validity)
- Pearson VUE — Claude Certification Program by Anthropic, used to confirm CCAR-F as the current exam code
- Microsoft Learn — AI-103, Azure AI Apps and Agents Developer Associate, the second credential in this programme
- The 66% compounding-reliability figure is plain arithmetic (0.95 to the power of 8), calculated by us to illustrate per-step error compounding
- Published Indian and US AI engineer and prompt/agentic specialist compensation ranges, used for role context only
360DT is an independent training provider. Certification exams are administered independently by Anthropic and Microsoft and are not included in programme tuition. Exam details are accurate as of 21 August 2026; always confirm current format, pricing, validity and blueprint version with the issuer before booking.