You can tokenize every decision, wrap your components in machine-readable contracts, and write specs an agent can't misread, and still watch the whole stack underperform for one mundane reason: none of it reached the agent at the moment it was working. Context that isn't delivered might as well not exist. That delivery problem is what the runtime files solve, and they're the least glamorous, most load-bearing layer of the whole tokens-to-agents argument.
Three files carry the pattern, each with a different job.
AGENTS.md carries architectural intent. Governed by the Agentic AI Foundation as an open standard, it's the always-on instruction file at the root of a repository: what this project is, the standing rules, the constraints that must never drift, and pointers to everything else. When an agent opens the repo, this is the first thing it reads. If your system has rules that are expensive to break (ours include "every visual value comes from a token" and "never use em dashes, anywhere"), this is where they live, because this is the file that's always in context.
DESIGN.md pairs machine-readable design data with human reasoning. Formalized by Google Stitch in March 2026, its move is the front matter: tokens in YAML at the top, prose underneath explaining the visual identity those tokens serve. One file, two audiences, no divergence, because the data and its explanation travel together. It's the same insight as the Indeed benchmark (structured data for contracts, prose for reasoning) folded into a single document.
SKILL.md packages capability. Pioneered by Anthropic, a skill is a folder with instructions and optionally scripts that an agent loads on demand, when the task calls for it, rather than carrying it always. Skills travel between projects, which makes them the reuse layer: how "the way we author tokens" or "the way we write component contracts" stops being tribal knowledge and becomes a portable artifact.
Notice the common tongue: Markdown. That's not nostalgia. It's what these models were trained on at colossal scale, so it's the format they read with the highest fidelity. The runtime layer works with the grain of the models instead of inventing a format they'd have to be taught.
And notice the architecture hiding in the three files: always-on versus on-demand. AGENTS.md is small and permanent; skills are heavy and loaded when needed. That split is context economics. An agent's context window is a budget, and a team that dumps everything into one giant always-on file pays for all of it on every task, mostly to bury the one rule that mattered. Lean standing rules, detail on demand: the same discipline as a good token architecture, applied to instructions.
This layer is also where I have the most skin. Our platform at a Fortune 250 utility (about 8,000 users) was built in large part with exactly this tooling: markdown context files and skills carrying a serious share of the load. When I say these files are how intent reaches the agent, I'm describing how we shipped, not predicting how someone might work someday. Of the four layers in the stack, this is the one I've leaned on hardest, and the one whose absence I'd feel first.
One concrete pattern shows why the layer matters so much: semantic role delivery. Your token system can encode that color.feedback.danger is the error color, but the agent picking colors at runtime still has to be told to prefer semantic tokens, never hit primitives from application code, and check the description field when unsure. Those three sentences in AGENTS.md are what make the semantic layer operational instead of theoretical. The runtime file is the bridge between your data and the agent's behavior; skip the bridge and the dumbest failure comes back: red.500 on an error state, chosen because it looked right.
Where to start, if your repo has none of this: write a one-page AGENTS.md today. What this project is, five standing rules you actually enforce, where the tokens live, where the specs live. Resist the urge to write ten pages; the discipline is the point. Add a DESIGN.md when visual identity decisions start repeating in code review. Extract a skill the second you paste the same instructions into a third session. You'll be surprised how quickly the files stop feeling like documentation and start feeling like part of the system, because that's what they are: the part that talks.
The free starter kit at tokenstoagents.ai/kit includes an AGENTS.md template with the standing-rules pattern, ready to adapt.