One atlas for you and your agents.
Field-tested patterns for designing agentic systems — written once, published twice: as prose you read, and as a payload your agents consult.
One private source of truth
Every pattern is written once, in a corpus that stays private. Nothing downstream is authored by hand, so there is no second copy to keep honest.
Gated, and stamped
Only what passes the gate is published, and what ships carries its provenance — so drift between the corpus and the public material is detectable rather than silent.
Two readers, one catalog
The slice is projected twice: a human edition to read, and an agent Catalog to consult. Neither is a summary of the other.
The Context Economy
The price model for the window: what an artifact costs to admit, in quality, dollars, and latency, and in whose window the bill lands.
context-economy#definition> atlas_cards(ids=["context-economy"], expected_revision="v1.0.2")
{
"projection": "agent-projection/v2",
"coherence": "v1.0.2",
"payload_class": "cards",
"data": { "cards": [{
"id": "context-economy",
"kind": "concept",
"claims": [{
"id": "window-price-model",
"text": "The price model for the window: what an artifact costs to admit, in quality, dollars, and latency, and in whose window the bill lands.",
"sources": ["context-economy#definition"]
}]
}]}
}
The two panels cannot drift apart, because a promotion swaps the Release and the Catalog derived from it in one transaction — there is no window in which the prose is newer than the payload. expected_revision is how a consultation stays pinned to one sealed publication: carry the revision a payload handed back, and a call that would cross a promotion is refused rather than quietly answered from a different Atlas. The response is shown abridged; the real one carries the card's other claims and the hints for recovering what this call left out.
Why it exists
Design knowledge written for people is the wrong shape for the agents doing the work, and knowledge written for agents is unreadable. The usual answer is to write it twice and let the two copies drift. Agentic Atlas writes it once and projects it twice, which makes the two editions the same fact rather than two descriptions of it.
Consultation is a ladder, so its cost stays proportional to relevance rather than to the size of the corpus. An orienting call returns identities and hooks; a card call returns claims for the few subjects that survived orientation; a full read is spent only on the one that turned out to matter. An agent that needs a name never pays for a chapter.
What the outside view exposed
The browser worked while ordinary machine clients failed: the MCP server rejected a default Accept header, and crawler HEAD requests returned an error. I reproduced scanner findings as real clients, fixed the protocol gaps, and made the API version a contract clients could request, rather than just a response stamp.
The result, and where I stopped
The ora readiness score rose from 73 to 88 out of 100 between August 24 and 27, 2026. These are point-in-time readings from that evaluation, not a permanent rating or a measure of adoption. The remaining meaningful gap was discovery beyond the site.
I declined recommendations for authentication, write operations, and collected ratings: the product has none of them. Making those absences explicit kept the public contract honest, even when adding them might have satisfied a scanner.
Read the evaluation, fixes, and tradeoffs →
Live since 18 August 2026. Built solo: writing, design, and engineering.