Kiro's central argument is that the problem with agentic coding is not the code generation, it is that nobody agreed what was being built before the generation started. The workflow it teaches has you turn a prompt into a requirements document, then a design, then a task list, and only then implementation, with a review gate at each stage. That is not a new idea in software engineering. It is a genuinely new idea in AI tooling, where the default is to prompt and hope, and the documentation explains it well enough that you understand why the structure exists rather than just how to click through it.
The requirements stage does more work than it looks like it should. Forcing the model to write down what it thinks you asked for, in a form you can read in two minutes, catches misunderstandings that would otherwise show up an hour later as three hundred lines of wrong code. I have watched this save real time. The design stage is similarly useful on anything touching more than one file, because it surfaces architectural choices while they are still cheap to change.
Steering files handle the problem every AI tool has, which is that you keep explaining the same conventions. Putting project standards, patterns and constraints in a file the agent reads every time is a better answer than hoping a memory feature retains what matters, and the documentation covers scoping them properly so you are not loading irrelevant context into every request. Hooks are covered well and the examples are realistic. Running tests on save, updating documentation when an interface changes, checking for secrets before commit, these are the things people actually want automated, and the material shows the mechanics without overselling the reliability.
The weight of the full workflow is the honest criticism. For a one line fix or a small refactor, going through requirements and design is ceremony, and the documentation presents the workflow as the way to work rather than as one mode among several. In practice you want the spec process for features and something much lighter for everything else, and you have to work that out yourself. Pricing is the usual mess.
Credit consumption varies by model and by how much the agent does, the relationship between what you build and what you pay is not obvious, and comparing it against competitors requires arithmetic the documentation does not do for you. Start on the free tier and watch the meter before committing. The AWS integration is thinner than the name suggests. This is a general purpose agentic IDE that happens to come from Amazon rather than a tool that gives you special leverage on AWS infrastructure, and anyone picking it up expecting deep cloud integration will be mildly disappointed.
Not a flaw, just a mismatch between branding and product. The gap I would most like closed is what to do when the spec itself is wrong. The workflow assumes that if you approve the requirements and design, the implementation will follow, and mostly it does. When the design was subtly mistaken, you have now spent tokens building the wrong thing carefully.
Guidance on spotting a bad design before approving it would be worth more than another page on hooks. Three point eight. The clearest thinking in this category about how to make agent generated code reviewable, taught properly, and marked down only for treating one workflow as universal and for the usual opacity around cost.