AWS shipped a markdown file to stop your agent hallucinating DynamoDB calls

AWS's own documentation now admits, in plain text, that coding agents "consistently get wrong" specific SDK calls — DynamoDB attribute marshalling in JavaScript, pagination in Python. The fix they shipped on May 6, 2026 is not a fine-tuned model or a smarter retrieval index. It is a markdown file. That gap between the problem and the shape of the fix is the whole story, and most of the "Skills vs MCP" takes flying around have it backwards.

What actually shipped

The Agent Toolkit for AWS is one package with three parts. There are 40-plus Agent Skills at launch. There is a fully managed AWS MCP Server with IAM-based guardrails, CloudWatch and CloudTrail observability, and sandboxed code execution. And there are three Agent Plugins — AWS Core, AWS Data Analytics, and AWS Agents for Bedrock AgentCore. It works with Claude Code, Codex, Cursor, and Kiro. It's free apart from the AWS resources your agent actually touches.

AWS's stated motivation is unglamorous and correct: agents "struggle with complex multi-service workflows, rely on outdated knowledge of AWS services," burning time and tokens and making teams reluctant to put agents near production. A companion post on dev.to says it without the corporate register — the agent "hallucinates AWS APIs because it's guessing from training data frozen in the past."

So: is this a new AWS product you have to adopt? No. It's a distribution channel for something SDKs have never had a home for.

It's SKILL.md, not an AWS format

Each AWS skill is a directory: a SKILL.md file, plus optional reference files, scripts, and slash commands. Loading one costs a few thousand tokens. This is not a proprietary format. It is Anthropic's SKILL.md spec, unchanged.

Anthropic launched Agent Skills on October 16, 2025 — a folder, a SKILL.md with YAML frontmatter carrying a required name and description, and bundled files alongside. The mechanism worth understanding is progressive disclosure across three levels. Level 1 is just the name and description, loaded at startup — one estimate puts the median at around 80 tokens per skill. Level 2 is the full SKILL.md body, loaded only once the model judges the skill relevant. Level 3 is the bundled reference files and scripts, pulled on demand. Anthropic's own analogy: a well-organized manual that opens with a table of contents, then specific chapters, then a detailed appendix.

On December 18, 2025 Anthropic released Skills as an open standard. Microsoft, across VS Code and Copilot, and OpenAI, across ChatGPT and the Codex CLI, reportedly shipped support inside 48 hours. OpenAI's Codex docs now say the quiet part out loud: "because Codex CLI uses the open SKILL.md standard, you're not locked into one agent — every skill works in Claude Code, Cursor, and Gemini CLI without modification." AWS is not opening a front in a format war. It is one more vendor writing to a format that already won.

Why not fix the docs, or use MCP, or use RAG?

This is the question that decides whether Skills matter, so it's worth being concrete about why each obvious alternative falls short.

Static docs go stale and bloat context. That's exactly the failure AWS names — training data frozen in the past, and docs the model has to be handed in full to be sure it read the right part.

MCP is the interesting comparison, because it's the one people frame as the rival. MCP, introduced by Anthropic on November 25, 2024, solves the N×M integration problem: every agent needing a custom connector to every tool and data source. It is a full protocol — hosts, clients, servers, resources, prompts, tools, sampling, three transports. Simon Willison's observation on the day Skills launched still holds: a skill, against all that, is just a Markdown file, which is "closer to the spirit of LLMs — throw in some text and let the model figure it out." But the sharper technical point is about where the cost lands. MCP tool schemas load as a flat, static list at every reasoning step. At scale that's permanent context weight. Skills' progressive disclosure keeps an unused capability nearly free — 80 tokens — until the moment it's invoked.

RAG retrieves prose. A skill ships a tested procedure. "Here is how you marshal a DynamoDB item, and here is the script that does it" is not the same artifact as three paragraphs of documentation about marshalling that happened to score well on a similarity search.

And MCP schemas encode what a tool is, not the right way to use it in sequence. That's the gap a skill fills. Which is why the war framing is wrong on the merits: AWS wires them together. Their docs state it directly — "Skills work best with the AWS MCP Server... Skills also work without it." The MCP server exposes search_documentation and retrieve_skill so an agent can discover skills live at runtime; the skills also install four other ways, including npx skills add aws/agent-toolkit-for-aws/skills. Both Anthropic and AWS treat the two as layers. Anthropic says Skills "complement" MCP by teaching workflows that involve external tools. The trade-press consensus lands in the same place: production setups use both.

The layering is easier to see once you drop the vendor logos. MCP is the wiring — how an agent reaches a tool. A skill is the operating manual — how to use it without making the mistake everyone makes. You don't argue over whether to keep the wiring or the manual.

The real fight is provenance

Here's what the toolkit is actually staking a claim on. Skill marketplaces have exploded: SkillsMP reports 2.2 million-plus indexed SKILL.md files, Vercel's skills.sh leaderboard lists 600,000-plus, LobeHub 300,000-plus. The honest commentary on all of it is that most are duplicated, abandoned, poorly documented, or never tested in a real workflow.

Against that backdrop, read AWS's pitch again. "Validated, tested," vendor-authored. The value isn't that AWS wrote smarter markdown than the community. It's that AWS controls the channel end to end — it authors the skill, tests it against its own SDKs, validates it, and ships it under its own name. When your agent pulls the DynamoDB skill, the provenance is the entire product. That's a platform moat forming in real time: official authorship as the thing you trust.

What to do about it this quarter

Two concrete moves.

First, audit which skills your agents actually load. In a 2.2-million-file marketplace, an agent can silently pull an unvetted skill and act on its procedure with full confidence. You want to know the difference between a skill your vendor shipped and one that surfaced from a leaderboard.

Second, treat unvetted community skills exactly like unvetted npm packages. A SKILL.md is executable intent — it can carry scripts and slash commands, and it directs what your agent does against your infrastructure. It is a supply-chain surface. Pin it, review it, or don't run it.

The format war is over; SKILL.md won, and it's multi-vendor. The war that's just starting is quieter and it's the one that matters: who you trust to write the SKILL.md — the vendor, the community, or the model's own frozen guess. Pick deliberately, because your agent won't.