The 9-Layer Agentic AI Stack
How modern organizations build, run, and secure AI agents at scale — and why every skipped layer becomes an incident.
Agentic AIArchitecture
Studies · Field Notes · Agentic AI · Guardrails
Systems that keep themselves honest, secure, and legal.
Research, architecture, and working notes on agentic AI — local-first, auditable, and built where the rules are real. Written from inside the build, where the failure modes actually live.
Regulated & security-first
Built where the rules are real
Security & compliance →Local-first by default — MAX3, REEF, and FeedHacker keep data on your own hardware, so there's no third-party egress of PII/PHI, the smallest possible breach surface, and a clean data-residency story. Automation stays approval-gated, reversible, and audited.
Track record
Delivery in high-stakes systems
$3.1B
multi-brand health plan modernized
~50%
operational efficiency gain from system unification
200%
platform performance lift (~$240K/yr saved)
10×
scale delivered at a Series A–C EHR/RCM startup
This site: HTTPS only · no third-party trackers · no analytics cookies · security.txt
Latest
From the field notes
All research →Classic RAG vs GraphRAG vs Agentic RAG
Choosing the right retrieval pattern for the shape of your knowledge — by the problem, not the hype cycle.
Agentic AIRetrieval
Claude Code Agent Teams
Production patterns for multi-agent collaboration: orchestration roles, review gates, and failure containment.
Agentic AIMulti-Agent
Build log
Selected projects
All projects →MAX3
A local-first, voice-first AI resident built as a single Python process — a DuckDB graph spine for memory, a Forward-Forward learner that predicts, proves and adjusts, and a voice loop to talk to him. The collective's flagship.
PythonVoice-firstFlagship
REEF
A local-first network-visibility and gateway-defense tool — a named fleet of sensors fused into one high-confidence card that sees it, explains it in plain language, and acts only on your approval.
PythonOPNsenseSelf-Defending
The Agentic AI Builder's Playbook
A 25-part published series on production-grade agentic AI architecture: the nine-layer agent stack, MCP and A2A protocols, RAG variants, agent memory, and evaluation & observability.
Writing25 Parts
Founder
About the founder
Full story →Jason Newell is the founder and principal architect of MAX Research Collective. He brings 25+ years of engineering leadership at the intersection of production agentic AI and regulated healthcare, and is the author of The Agentic AI Builder's Playbook, a 25-part series on production-grade agentic AI architecture. His current research centers on MAX3, a local-first, voice-first AI resident, and the theory behind it.