Ten whitepapers.
Zero hand-waving.
AutoArchitect's architecture, documented as focused whitepapers — the substrate, the single-Architect runtime, the 25+ tool palette, the plugin catalog, the connection vault, the cron scheduler, the knowledge/RAG store, the preview pipeline, and what's next on the roadmap. Each opens with an Abstract + Why-this-exists, dives into mechanics with concrete code references, then names the failure modes the design handles.
Read in order if you're new. Skip directly to the section that matters if you're integrating. The whitepapers are the source of truth — the README and this site are summaries of them, not the other way around.
Architecture Overview
The map of all maps. Component diagram, principal flow (one prompt → deployed app), data flow per primary entity, repository layout, performance + cost characteristics, what's intentionally not in v50.
Read on GitHub →The Substrate
autoarchitect-code in detail. Library +
CLI + roadmap-HTTP-service. The 27 tools, the 12 safety
invariants, the 5-level permission gate. The 8-step
gap-fill loop that authors a fresh skill or archetype.
How the platform invokes it deterministically.
The Agent Fleet
Why 10+1 instead of 25+. Anatomy of an agent (SOUL + cron + tools + memory). Coordination mechanisms (task graph, group-chat, typed bus, direct delegation, shared memory). The work-loop. Task-kind postconditions. Dynamic spawning recipe.
Read on GitHub →Skills + Archetypes
Skill shape. Catalog organization (connectors,
templates, automations, domains). Archetype anatomy.
matchTemplates algorithm. Why skills, not
MCP. Runtime catalog vs baked. Promotion workflow.
Task Contracts
The Task table schema. The 15 task kinds + their postconditions. Deliverable lifecycle. Task readiness gates (deps + qaReady + PAC non-empty). Dependency graph correctness. Lease + retry semantics. Dead-letter handling.
Read on GitHub →The Verification Pipeline
PAC schema. Verification kinds (preview, mount, data,
functional, autonomous, acceptance, citation, length).
QA verdict templates. qaReady runtime gate.
Critic vs QA. Why PM owns the PAC table single-source-
of-truth.
Sensorium · Invariants · Oracle
The runtime observable to itself. Sensorium snapshot + typed queries. InvariantRegistry + 30s HealthTick. The Oracle remediation reasoner. TypedBus pub/sub. Dashboard surfaces.
Read on GitHub →The Runtime Guards Philosophy
SOULs are aspirational. Runtimes are mandatory. The single architectural insight at the core of v48–v50, learned the hard way across seven E2E runs that burned through tokens for nothing. Every load-bearing rule must be enforced by a deterministic runtime guard, not by hopeful language in an agent's system prompt. This whitepaper enumerates the guard taxonomy, when NOT to guard, and the harness-as-contract pattern that makes the rule survive every code change.
Read on GitHub →The Self-Healing Pipeline
Liveness Probe (every 30s). Smoke-worker (headless Chromium service). Conductor's wakeJudge → fix-task routing. The v48 backoff: per-slug counting + escalation to substrate after 3 + dead-letter sweep of pending Builder fix tasks.
Read on GitHub →Roadmap
v50.1 fast-follow patches. v51 operator quality of life (catalog promotion automation, fleet templates, budget alarms). v52 ecosystem (MCP wrapper, marketplace, LLM routing, federated authoring). v53+ multi-tenant cloud. Open architectural questions. What we're explicitly NOT doing.
Read on GitHub →Three reading paths.
Linear (00→09)
If you're new and want the full picture. ~2,700 lines total, ~90 minutes of reading. Each builds on the prior.
Substrate-first (00 → 01 → 07 → 02)
If you want to know "what makes this different from other multi-agent frameworks." The substrate + the runtime guards philosophy + the agent fleet's design choices. ~45 minutes.
Operator path (00 → 04 → 05 → 08)
If you're running AutoArchitect in production and need to understand task contracts, the verification pipeline, and the self-healing path. ~60 minutes.