Decompose Agentic System

Architect recursive system and task decomposition for large-scale multi-agent architectures (100+ files, 50K+ tokens), DAG dependency trees, disk-backed persistent planning, and policy-as-code fitness functions.

Core Rules

• Recursive Work Breakdown: Decompose massive workloads (100+ files, 50K+ tokens) hierarchically into bounded sub-agent scopes; prohibit monolithic single-agent execution on massive codebases.
• Hierarchical DAG Dependency Trees: Structure execution as Directed Acyclic Graphs (DAGs); establish topological sorting and strictly prohibit unconstrained cyclic agent loops.
• Invariant Gating: Enforce automated pre-condition and post-condition assertions at every phase boundary; halt execution immediately on invariant failure (AGENTIC-DECOMPOSITION LOCK).
• Disk-Backed Persistent Planning (planning-with-files): Maintain execution plans, task manifests, and progress logs in durable markdown files on disk (plan.md, task.md, progress.md, BRIEFING.md); never store long-horizon state in ephemeral context (CONTEXT-ROT-RESISTANCE LOCK).
• Deterministic Completion Verification: Require empirical verification evidence (test outputs, compiler exit code 0, schema validation) before task closure; reject LLM self-declaration of completion.
• Anti-Context-Rot Discipline: Enforce periodic liveness updates (Last visited: [ISO-8601]), smart zone prompt budgeting, and compact briefings (< 100 lines) surviving context compaction and session resets (/clear).
• Non-Human Identity (NHI) Autonomy Tiers: Classify sub-agent tasks into Tier 1 (Supervised / HITL), Tier 2 (Semi-Autonomous / Guardrailed), or Tier 3 (Fully Autonomous / Sandboxed).
• Blast-Radius Perimeter Isolation: Confine sub-agent filesystem access, network egress, and token capabilities to designated module directories and sandboxes (BLAST-RADIUS LOCK).
• Policy-as-Code Architectural Fitness Functions: Codify architectural constraints into automated CI fitness functions using Cedar (agent trust boundaries), OPA/Rego (module boundaries), or calibrated LLM-as-judge (FITNESS-FUNCTION LOCK).

When to Use