Highlighted ADRs

Of the 33 ADRs the platform has written, six are foundational — they shaped what the platform IS and how every other decision relates to them. Each one gets the docs-site treatment here: the original ADR, plus narrative context for someone landing on it fresh.

If you read these six in order, you have the platform’s architectural worldview.

Identity: what the platform IS

ADR-0005 — Two-layer architecture Why the platform splits into Platform Core (business-agnostic) and Business Packs (vertical-specific). The decision that determines what 'agent platform' even means.

ADR-0006 — Six-layer context system The platform's most distinctive design choice. Six strictly ordered context layers; layers 1–2 immutable. The security boundary against prompt injection lives here.

Engine: how agents run

ADR-0021 — Agent runtime and tool loop The execution engine. Tool loop, budget enforcement, deterministic stop conditions. How an LLM call becomes an agent turn.

ADR-0022 — Delegation as tool The unification trick. Sub-agent delegation is implemented as a tool call, not a separate API. Why this single decision keeps the runtime small.

Distinctive: what makes Phase 1 hard

ADR-0030 — Long-term memory access pattern The hardest design problem in Phase 1. How agents read and write persistent memory; per-tenant scoping; embedding-on-write; the interaction between Vectorize, D1, and the embedding adapter.

ADR-0031 — YAML agent definition format What an agent IS in this system. The schema, the loader, why YAML beat the alternatives, what gets validated where.

Why these six

Every ADR matters; these six have an outsized share of the platform’s identity. Read them and you can answer:

• What is the platform? (0005, 0006)
• How do agents run? (0021, 0022)
• How do agents remember? (0030)
• How are agents authored? (0031)

Other ADRs answer follow-up questions — error taxonomy, model adapters, deployment, jobs. Important, but downstream of these six.

The full list of 33 is on the All ADRs index.