A unified architecture for production-grade multi-agent systems with phase-aware orchestration and persistent knowledge compounding. Four components. A five-phase taxonomy. Running in production today.
Multi-agent AI systems suffer from four fundamental problems: context loss between sessions (agents restart from scratch), naive routing (keyword matching instead of phase-aware allocation), brittle retrieval (single-pass search without credibility control), and raw prompting (no compilation, no parameterization, no reproducibility).
ROSTR solves all four. It formalizes agent invocation into a compiled manifest (PAL), retrieves knowledge with stratified credibility (RAG DAL), classifies and allocates work with a 5D phase taxonomy and 4D priority score (NPAO), and persists state across sessions with scoped namespaces (Rostr Hub).
The framework was published in April 2026 through Zenodo with DOI 10.5281/zenodo.19550414. It is the engine behind the Gency AI agency — turning SOPs into intelligent agents — and a contribution to the multi-agent systems field.
PAL transforms loose natural language into strict agent runtime manifests. Instead of sending raw prompts to models, every agent invocation flows through a five-stage compilation pipeline:
RAG DAL replaces naive single-pass search with a three-tier credibility architecture:
The retrieval algorithm runs multi-pass: broad sweep → gap fill → deep verification. Each pass assesses coverage per sub-topic using a confidence scoring formula that weights source count (35%), consistency (30%), tier distribution (25%), and recency (10%). Pass 4 only triggers if two or more topics remain below the 0.6 threshold — at which point they're marked unresolvable.
NPAO is the operational brain. It classifies every task into one of five phases, scores it on four dimensions, and allocates it to the optimal agent.
5D Phase Taxonomy:
4D Priority Scoring: Phase Urgency (0-10) × 0.35 + Dependency Impact (0-10) × 0.30 + Business Impact (0-10) × 0.25 + Resource Efficiency (0-10) × 0.10. Tasks scoring ≥7.0 get immediate allocation; 4.0-6.9 are queued; below 4.0 goes to backlog.
Agent Allocation: Eligible agents are filtered by phase compatibility, tool availability, and dependency status. Each is scored on context similarity (50%), specialization overlap (35%), and current load (15%). Highest scorer wins.
Orchestration patterns: Sequential chain (A→B→C), parallel fan-out (A→[B,C,D]), aggregation fan-in ([A,B,C]→E), conditional branch (A→decision→B or C).
The Hub is what makes knowledge compound across sessions instead of evaporating. Four levels of state:
Cross-namespace access requires permission. Knowledge retrieval uses hierarchical credibility — not all sources are equal. Every project gets a directory under rostr-hub/projects/{id}/ with knowledge-base, learnings.jsonl, timeline.jsonl, and checkpoints.
The ROSTR framework powers the project intake system at Gency AI. When a client submits a raw SOP or idea — notes, files, links, a one-line description — PAL compiles the intent, NPAO classifies the work phase and scores priority, and the system outputs a complete project package: knowledge base, PRD, architecture diagram, reporting framework, and phased build plan. Idea → fully-scoped agent: hours → under 30 minutes.
The agency's core product pipeline runs on ROSTR: JTBD Analysis captures the operator's functional requirements → PAL/NLP extracts and compiles the intent → the ROSTR System delivers the agent, workflow, or skill. Every client engagement from SOP to deployed agent follows this path.
The data sync pipeline that ingests 1,500+ business licenses from Chicago Data Portal daily uses ROSTR's NPAO task router for phase classification and cron orchestration. $0/month infrastructure on Vercel free tier.