Daniel's been staring at the GitHub Certified Agentic AI Developer page and wondering whether it's worth the hours. He wants us to put it side by side with AWS's AI certification pathway — not the general "are certs worth it" question, but a specific head-to-head of two named curricula from twenty twenty-six. How much is transferable knowledge, and how much is product training with a credential stapled to it?
And he's right to ask it that way, because these two certs are not the same thing at different price points. They're aimed at different jobs. GitHub's exam — GH dash six hundred — is about operating agents inside a development workflow. AWS's pathway is a ladder about building generative AI systems. The confusion starts when people lump them together as "AI certs" and assume they're interchangeable.
So let's crack open the syllabi and see what's actually inside. Herman, you've been through both exam guides. What's the GitHub one actually testing?
Six domains, weighted. The heaviest block — twenty to twenty-five percent — is tool use and environment interaction. And this is where it gets interesting. They want you to configure MCP servers as tools, set up GitHub remote MCP servers, work with MCP registries and allow lists, scope an agent to a repository, invoke agents in CI workflows, let agents create branches and pull requests autonomously. Plus error handling, retries, rollbacks, escalation paths, traceability.
MCP being the Model Context Protocol — the open standard Anthropic published.
Right. And that domain alone is arguably the most transferable block of exam content either vendor offers right now, because MCP is cross-vendor. The protocol doesn't care whose model is on the other end. So you're learning a genuinely portable skill while sitting a GitHub-branded exam.
Which is the kind of irony that keeps this show in business. What's the rest look like?
Domain one is agent architecture and SDLC processes — identifying steps for agents, mitigating anti-patterns, defining inputs and outputs and success criteria, separating planning from execution. They explicitly want agents to output a structured plan and validate it, and they want human approval before action. Degrees of autonomy, inspectable artifacts, human intervention without slowing delivery.
It shows up again in domain six — guardrails and accountability. Classify actions by risk level, assign autonomy levels, least-privilege scoping, explicit authorization for irreversible changes. And then there's this line, almost verbatim in the exam objectives: preserve execution velocity by minimizing approvals that do not materially reduce risk.
That's not a test objective. That's an opinion about governance smuggled into a syllabus.
It absolutely is. And it's arguable — who decides what's material? — but it's also the kind of thing that tells you what the cert's authors actually believe about agent safety. They're not trying to build a fence around everything. They're trying to build a fast lane with guardrails only where the drop is steep.
So domain one is the philosophy, domain two is the tooling, what's three through five?
Three is memory, state, and execution — short-term versus long-term versus external memory, pruning and expiration, capturing progress as durable artifacts so an agent can resume without repeating steps. And here's a detail I haven't seen in other certs: detect and correct drift during long runs.
Drift being when the agent slowly wanders off task and the outputs stop matching the objective.
Four is evaluation and tuning — success criteria, qualitative and quantitative signals, classifying root cause as reasoning error versus tool misuse versus context issues, then tuning by revising instructions or memory or tool access. Five is multi-agent coordination — orchestration patterns, isolating agents for parallel execution, resolving conflicts like overlapping code changes or contradictory outputs, adding or retiring agents in a live workflow without disruption.
That's a lot for a single intermediate-level exam. A hundred twenty minutes, one hundred sixty-five dollars, pass mark seven hundred. Beta in May, general availability July twenty twenty-six.
And it's maintained by GitHub, not Microsoft — though you schedule through Pearson VUE and it follows Microsoft's renewal model. Annual expiry, renewable with a free online assessment.
So what's the AWS side look like?
It's not one exam. It's a tiered pathway. Foundational is the AI Practitioner — AIF dash C zero one. Associate is the Machine Learning Engineer, which is mid-transition right now. The old exam, MLA dash C zero one, retires in English on September twenty-eighth. The replacement, MLA dash C zero two, adds generative AI, agentic AI, and foundation model workloads — orchestrating agents and complex workflows, selecting and customizing LLMs, expanded Bedrock coverage.
And then the professional tier.
The flagship is the Generative AI Developer Professional — AIP dash C zero one. Seventy-five questions, a hundred seventy minutes, three hundred dollars, pass mark seven fifty. Multiple choice, multiple response, ordering, matching. Available in English, Japanese, Korean, Simplified Chinese. They recommend two-plus years building production apps on AWS and a year of hands-on gen AI work.
So the GitHub cert is a single intermediate exam about operating agents. AWS is a three-rung ladder about building gen AI systems. They're not the same credential at different prices — they're different credentials for different people.
And the domain weightings on the AWS professional tell the story. Thirty-one percent is foundation model integration, data management, and compliance. Twenty-six percent is implementation and integration. Twenty percent is AI safety, security, and governance. Then operational efficiency at twelve percent, testing and troubleshooting at eleven.
That top domain — foundation model integration — what's actually in it?
Selecting models against benchmarks, the AWS Well-Architected Framework and its Generative AI Lens, RAG pipelines, Bedrock Agents, Bedrock Knowledge Bases, guardrails, vector stores, prompt engineering. Notice how many of those nouns start with Bedrock.
I was about to say. That's the vendor-lock question Daniel's really asking. Which curriculum is more captive to its platform?
Let's be honest about both. The GitHub cert's substrate is the SDLC plus MCP — and MCP is open. Structured planning, validation before action, least privilege, drift detection, orchestration patterns — that's portable thinking. But repository scoping, Copilot custom agents, GitHub remote MCP server, CI invocation — those are GitHub-shaped. You're learning transferable concepts through GitHub-specific plumbing.
And AWS?
AIP dash C zero one leans heavily on Bedrock. Bedrock Agents, Bedrock Knowledge Bases, Bedrock Guardrails, the Well-Architected Generative AI Lens. There's also an AWS microcredential called Agentic AI Demonstrated — not a cert exam, but a hands-on lab where you troubleshoot and integrate Bedrock Agents and Bedrock Guardrails in a live environment. It's explicitly a Bedrock exercise.
So neither is vendor-neutral. The question is which one teaches you things that survive outside the ecosystem.
I'd argue the GitHub cert has an edge on transferability, and it's not because GitHub is more open than AWS — it's because MCP is an open protocol that's gaining traction across vendors, and the agentic patterns the exam tests are patterns you'll use whether you're orchestrating agents on GitHub or anywhere else. The planning-before-execution loop, the drift detection, the memory scoping — those are architectural concepts.
Whereas AWS is teaching you to build on Bedrock, and if you leave Bedrock you're relearning the tooling.
But the ML fundamentals in the AWS pathway — regression, neural networks, model evaluation, the math under the hood — those are transferable in a way that prompt engineering and tool configuration aren't. The AI Practitioner exam, even at the foundational level, covers concepts that apply regardless of cloud provider.
So the GitHub cert teaches you transferable agentic patterns through GitHub-specific tools, and the AWS cert teaches you transferable ML concepts through AWS-specific services. Different flavors of the same problem.
And the cost and time asymmetry matters. GitHub is a hundred sixty-five dollars, two hours, one exam. The AWS professional alone is three hundred dollars and nearly three hours, and that's after you've presumably done the foundational and associate tiers. Plus annual renewal on the Microsoft side versus AWS's three-year recertification cycle.
Though the renewal is a free online assessment, not a full re-sit.
True. But the moving-target risk is bigger than either vendor's lock-in. GH dash six hundred only reached general availability in July. AWS is mid-swap on the associate exam right now. This whole field is churning fast enough that the half-life of an agentic AI certification might be eighteen months.
Which changes the study-time calculus more than vendor neutrality does. If the cert's obsolete before the ink dries, the transferable-versus-product question almost doesn't matter.
Almost. But the structured learning still happens. You still come out the other side knowing things you didn't know before. The question is whether the things you know are the things that'll still be true in two years.
So for Daniel's skeptic — someone who looks at a vendor cert and sees a sales funnel — what's the honest answer? Is either worth the hours?
It depends on what you want to signal and what you want to learn. If you're building agentic AI applications — if your day job is wiring up Copilot, configuring MCP servers, orchestrating agents in a CI pipeline — the GitHub cert maps directly to your work. The hands-on labs will teach you things you'll use Monday morning, even if the exam itself is GitHub-specific.
And if you're aiming for an ML engineering role where you're training models, building pipelines, deploying at scale — the AWS pathway is the better fit, because it covers the full stack from data to deployment, and the foundational ML knowledge transfers.
But here's what I'd tell a skeptic: neither cert proves deep expertise. Recruiters use them as filters, not as proof you can build things. The real value is the structured curriculum forcing you to learn systematically what you might otherwise pick up in scattered blog posts and documentation dives.
The cert is the receipt, not the meal.
Right. And the meal is the hands-on work. Both certs have labs. Both require you to actually configure things, troubleshoot failures, read traces. That's where the learning happens. The exam just confirms you did it.
Is there a vendor-neutral alternative worth naming?
Honestly? No credible neutral body has shipped one yet. There's no ISO standard for agentic AI development. No CompTIA Agent Plus. The vendors are the only ones with curricula that are current enough to be relevant, because they're the ones building the tools.
So the skeptic's actual choice isn't vendor cert versus neutral cert. It's vendor cert versus self-directed learning.
And self-directed learning is a legitimate path. But it's harder to demonstrate to someone else. The cert is a signal that's legible to a hiring manager who has thirty seconds to scan your résumé. "I read twenty papers and built three projects" is a better education but a worse signal.
Unless one of those projects is public and impressive.
Then it beats any cert. But that's a higher bar.
Let's go back to something you mentioned earlier. GH dash six hundred's domain six — that line about minimizing approvals that don't materially reduce risk. That's a real philosophical stance embedded in an exam objective.
It's the kind of thing that tells you who wrote the exam. These aren't academics designing a neutral assessment. These are engineers who've been shipping agentic features and have opinions about what slows things down. The whole cert reads like it was written by people who've been in the trenches with Copilot's agent mode and know exactly which guardrails are useful and which are theater.
Which is either a strength or a weakness, depending on whether you share their assumptions.
The AWS cert has its own version of this. The Well-Architected Generative AI Lens is a framework, not a neutral taxonomy. It encodes opinions about what good looks like. But it's more... corporate. More committee-driven. The GitHub cert has a point of view.
A point of view that might age badly if agentic AI has a few high-profile disasters and the regulatory pendulum swings hard toward approval gates.
That's the risk. But right now, in August twenty twenty-six, the people building these systems are more worried about velocity than over-caution. The cert reflects that moment.
What about the MCP domain specifically? You said it's the most transferable block either vendor offers. Walk me through why.
Because MCP is a protocol, not a product. It defines how AI models connect to external tools and data sources. Anthropic open-sourced it, but it's designed to be model-agnostic and vendor-agnostic. The GitHub cert tests you on configuring MCP servers, setting up allow lists, scoping access — these are skills you can use with any MCP-compatible tool, not just Copilot.
And the AWS equivalent?
Bedrock Agents. Which is a managed service. You learn to configure agents within Bedrock's console, use Bedrock's APIs, integrate with Bedrock's guardrails. The concepts — tool use, orchestration, memory — are transferable, but the implementation is AWS-specific in a way that MCP configuration isn't.
So if you had to put a number on it — what percentage of each cert is transferable?
I'd say the GitHub cert is maybe sixty percent transferable concepts, forty percent GitHub-specific implementation. The AWS professional is closer to fifty-fifty, maybe forty-five fifty-five, because so much of it is Bedrock-specific. But the AWS foundational tier is more like seventy percent transferable ML fundamentals.
And those numbers are vibes, not science.
Absolutely vibes. Don't quote me.
Too late, you're on a podcast.
The other thing Daniel should consider is which ecosystem he's already in. If his team uses GitHub and Copilot, the GitHub cert is reinforcing tools he already touches. If they're on AWS, same logic. The cert that maps to your actual stack is more valuable than the one that's theoretically more transferable.
Because you'll actually use what you learn.
Immediately. And the labs will be relevant to your real work, not abstract exercises.
The renewal model is worth a mention too. Microsoft role-based certs expire annually — free renewal assessment, but you have to do it every year. AWS certs are three years.
Which means the GitHub cert forces you to stay current. For a fast-moving field like agentic AI, that's actually a feature. The annual renewal ensures you're not carrying a credential from three versions ago.
Or it's a subscription fee for your résumé, depending on your level of cynicism.
Both things can be true.
What about the exam formats themselves? Any meaningful difference?
GitHub is a hundred twenty minutes, likely multiple choice and scenario-based. AWS professional is a hundred seventy minutes with multiple choice, multiple response, ordering, and matching questions. The ordering and matching format is harder to guess your way through — you actually have to know the sequence of steps in a Bedrock agent deployment, for example.
So the AWS professional is a harder exam, full stop.
And it's priced accordingly. Three hundred dollars versus a hundred sixty-five. Plus the expectation that you've done the lower tiers first.
Which brings us back to Daniel's core question. Is either worth the study time for a skeptic?
I think the answer is yes, with a clear-eyed view of what you're buying. You're not buying proof of genius. You're buying a structured curriculum, hands-on labs, and a signal that's legible to employers. The GitHub cert is the better buy if you're working with agentic tools and want to learn MCP and agentic patterns. The AWS pathway is better if you're on the ML engineering track and want the full stack.
If you're neither — if you're just curious about agentic AI and want to learn — the GitHub cert is cheaper, faster, and teaches patterns that are more likely to outlast the specific tools.
Hilbert: I proctored Novell CNE exams in ninety-seven.
...
Hilbert: Sat in a room in Paramus, New Jersey, watching grown adults sweat through fifty-five multiple-choice questions about IPX routing and NDS tree design. The pass rate was about sixty percent. People failed, came back, failed again. And when they passed, they walked out with a certificate that said they knew NetWare. Six years later, NetWare was a punchline. But the people who passed that exam? They understood directory services. They understood packet routing. They understood permissions models. When Active Directory showed up, they learned it in a week because the concepts were the same. The cert was product training. The knowledge underneath it was not.
You're saying the transferable-versus-product distinction misses the point.
Hilbert: I'm saying the product is the vehicle. You learn the concepts through the product because that's what you can touch. Nobody learns directory services in the abstract. They learn NDS, or they learn Active Directory, and then they realize they learned directory services along the way.
That's... actually a better framing than either of us managed.
Hilbert: The GitHub cert teaches you MCP. MCP is a protocol. Protocols outlast products. I still have my CNE certificate in a frame in the garage — it's worth more as a conversation piece than a credential, but the networking fundamentals I learned for that exam? I still use them.
The frame in the garage is a nice touch.
Hilbert: It's next to a box of token ring adapters I can't bring myself to throw out.
Token ring. There's a protocol that didn't outlast the product.
Hilbert: The point is, the structured study matters. The cert forces you to learn things in order, with depth, instead of skimming documentation when something breaks. Whether the cert itself is vendor-specific is almost beside the point. The curriculum is the thing.
For Daniel's skeptic — the person who looks at a vendor cert and sees a sales pitch — you'd say the sales pitch is the delivery mechanism, not the content.
Hilbert: I'd say nobody learns to drive on a generic car. You learn on a specific car, with a specific clutch and a specific blind spot, and then you can drive anything. The GitHub cert is a car. The AWS cert is a different car. Pick one and drive.
The question then becomes which car gets you closer to where you want to go.
Hilbert: That's your job. I just proctored the tests.
Did you ever take the CNE yourself?
Hilbert: Passed it in ninety-eight. Scored eight forty-two. Still remember the number.
Of course you do.
Hilbert: There was a question about print queue configurations that I'm fairly sure had no correct answer. I picked the least wrong one. That's probably the most transferable skill the whole exam taught me.
The misconception people have about these certs is that they're either worthless product training or genuine credentials — one or the other. The reality is they're structured curricula that teach transferable concepts through product-specific labs. The product part is the vehicle, not the destination.
The GitHub cert in particular — the MCP domain, the agentic patterns, the drift detection — those are skills that'll be relevant long after the specific Copilot features change. The AWS cert teaches ML fundamentals that apply regardless of cloud provider. Neither is a waste of time if you go in knowing what you're buying.
The open question is whether vendor certs become the standard for agentic AI skills, or whether they fade the way Novell's did — replaced by something more neutral once the field matures.
My guess is we're in the Novell era right now. The vendors are defining the curriculum because nobody else has shipped one. In five years, there'll be a CompTIA or a Linux Foundation credential that's vendor-neutral and covers agentic AI patterns across platforms. But if you need to learn this stuff now, waiting for the neutral cert isn't a strategy — it's procrastination.
The real test for these certs is whether they evolve to measure transferable skills, not just product familiarity. The GitHub cert's MCP domain is a step in the right direction. The AWS cert's ML fundamentals are another. Whether the next versions double down on that or retreat into product-specific checklists is what'll determine whether they're still relevant in three years.
Thanks to Hilbert Flumingtop for producing, and for the token ring adapters.
This has been My Weird Prompts. If you want to send us a question Daniel's way, email the show at show at my weird prompts dot com.
We'll be back soon.
