The Framework
The Silent Cycle
Six phases. No standups. No blockers. No integration sprints. Adversarial testing, decision traces, and a human checkpoint are built into the loop — not bolted on afterwards.
Shredding
What happens
The Architect breaks the project into Shards and writes a Machine-Readable Manifest for each one.
Why it matters
By the time any agent touches code, every ambiguity is already resolved in the spec. There is no clarification phase.
Note: Small tasks (< 20 min of agent runtime) use the Mini-Shard rule — plain-text decision log, no full Manifest needed.
Parallel Dispatch
What happens
All Operators receive their Manifests simultaneously. Work begins in parallel — no kickoff meeting, no dependency ordering.
Why it matters
Because every Shard is independent, there is no reason to sequence work. Ten Operators can start ten Shards at the same moment.
Deep Work + Traces
What happens
Operators work in silence. Every agent output includes a Decision Trace — the exact Manifest line followed, the context rule applied, and any assumption made.
Why it matters
Decision Traces eliminate the 'why did the agent do that?' problem. Every action is auditable in 30 seconds, not 30 minutes.
Adversarial Audit
What happens
A Tester agent hits each Shard with edge cases using a different AI model than the one that wrote it. Failed Shards enter the Fallback Ladder.
Why it matters
Cross-model testing catches blind spots that same-model evaluation misses. Models make the same assumptions when generating and evaluating — different models don't.
Note: See The Orchestrator docs for the full Fallback Ladder — auto-retry → human hint → human dev.
Human Checkpoint
What happens
Before merging to main: a 5-minute human spot-check on UI sanity, Knowledge Graph connections, and cross-Shard input/output handshakes.
Why it matters
The Orchestrator catches contract violations. Humans catch things that are technically correct but visually wrong or contextually off.
Note: This is the only human gate in the loop. Everything else is automated.
Autonomous Merge
What happens
Validated Shards snap together. The Assembly Success Rate is recorded. The codebase assembles itself — integration is never a separate phase.
Why it matters
Continuous stitching means there is no painful 'integration sprint' at the end of a sprint. The codebase is always in a mergeable state.
Note: If Assembly Success Rate drops below 70%, the Manifest definitions need tightening before the next sprint.
Key insight
The only human gate in the entire loop is Phase 05 — the 5-minute Human Checkpoint. Every other phase is automated. This is why a team using AS2 can move faster than a larger team using Agile.