Agentic AI · Guardrails · Evaluation
Systems that act, and the evidence they leave behind
The largest body of work on this site. Agents that plan, call tools and change things are not chatbots with extra steps — they fail differently, they cost differently, and they need evidence a human can audit after the fact. What follows is what building and breaking them actually taught.
An agent is not judged by what it can do — it is judged by what it did, and whether you can prove it.
Every demo works. The interesting question starts afterwards: what happened when the tool call failed, what the agent did with a stale document, who approved the write, and what the run cost. The writing here treats an agent as a system with a ledger rather than a model with a prompt — because that is the difference between something you can ship into a regulated environment and something you can only show people.
The frame
Four things an agentic system owes you
Read the stack →Intent
What was the agent asked to do, in words that survive being read back a month later? An instruction nobody can reconstruct is not a specification.
Action
What did it actually call, with what arguments, and what came back? The gap between intended and performed is where every interesting failure lives.
Evidence
Can somebody who was not there reconstruct the run? Logs that record success are decoration; logs that record the decision are evidence.
Reversal
What happens when it is wrong — approval gates, dry runs, and an undo that was designed rather than discovered during the incident.
Flagships
The stack, and the role that owns it
The 9-Layer Agentic AI Stack
How production agentic systems are actually assembled, layer by layer, with what each layer is responsible for and what breaks when it is skipped. The opening part of the Playbook, written for people building rather than evaluating.
The Case for an AI Resident
A role that did not exist until the work demanded it: someone who lives inside the system, keeps its memory, and is accountable for what it does. An argument about org design first and tooling second.
Why it fits
Why this is the biggest grouping here
Built, not surveyed
MAX3, REEF and FeedHacker are running systems, and most of this writing is a post-mortem of something that happened in one of them rather than a reading of somebody else's paper.
The failures are the content
A guardrail that never fired proves nothing. The pieces that matter here are the ones where something got through — and what was changed so it could not happen twice.
Evaluation over demonstration
An agent that passes a demo and fails an evaluation is the normal case. The writing keeps returning to evaluations that can actually fail, because those are the only ones worth running.
Regulated by default
These systems were built where PHI and PII are present, so local-first, approval-gated and auditable are not features added late — they are the starting constraints.
Field notes
Agentic AI writing
See all →- Technology Responsibility: A Compliance and Security StudyData
- NIST and Trustworthy AI: The Lab That Never Fines AnyoneAI Governance & Standards
- LangChain and LangGraph: What They Are, How They Work, and Why MAX Took Another RoadAgent Frameworks
- FedRAMP and GovRAMP: Evidence as a CommandGovernment Cloud Authorization
- DICOM and PACS on MAX4: Pictures Before PatientsMedical Imaging & De-identification
- Context Windows and Code Graphs: The Error Past the CutAI Code Review & Context
- Virtual Cell Models: The Bottleneck Is the DataHealthcare AI & Biology Data
- Reading Healthcare-AI Evidence: One Arm Is Not EvidenceHealthcare AI Evidence
- AI and Healthcare Payment: The $942 Million Nobody CheckedHealthcare AI & Payment
- Local-First Medicine: The Box in the ClinicLocal-First Clinical AI
- Patient Records and AI Agents: The Patient Holds the KeysPatient-Gated Health Records
- AI Voice Agents in Medical Practices: The 81% ProblemHealthcare Voice AI
- Clinicians and AI: What They Actually AskHealthcare AI Evidence
- AI Agent Liability: Who Goes to Prison When the Agent Breaks In?AI Agent Liability
- MARA Demo Update: Earned Authority and a Person on CallHealthcare AI Design
- Agile and Iterative Development: A Slice Through Every LayerSoftware Delivery & Engineering Practice
- MARA, the Medical Agentic Resident Assistant: Four Generations, One LineHealthcare AI Design
- MAX4 and MARA: The Ladder Starts at OffMAX4, Trust & Clinical AI
- What Building With AI Really Costs: The Whole BillAI Toolchain Costs & CI
- MAX: The AI Resident, and the Road to MARALocal-First AI & Healthcare
- MARA: A Medical Agentic Resident Assistant That Learns From Every Encounter and Never Makes the DecisionHealthcare AI Design
- Root Cause Analysis: The Check Name GitHub Never RenderedRoot Cause Analysis
- REEF, Local-First Network Defence: The Porch Camera's HeartbeatSecurity & Compliance
- An Engineering Organization Run by AI AgentsAgentic AI
- Hosted Code Review: The YAML Parsed and Nothing Read ItAI Code Review
- CapEx, OpEx, and EBITDA: The Same Dollar, Two Different StoriesTechnology Finance & Capital Strategy
- Depreciation and Useful Life: The Chip Outlives Its EconomicsInfrastructure Depreciation & Useful Life
- AI Spend: Three Budgets Pretending to Be OneAI Cost & Procurement
- AI and Headcount: Somebody on the Board Did the ArithmeticAI, Productivity & Headcount
- Data Pipelines for Machine Learning: The Duplicate That Hid the HeartbeatData Engineering & ML Pipelines
- Local-First Code Review: The Guard Dog and the LedgerLocal-First AI & Code Review
- Graph Engineering: The Edge That Carried NothingAgentic AI & Orchestration
- SAST and SCA: Nothing Matched the Patterns We Know AboutSecurity Scanning & Static Analysis
- Security Scanning in Practice: The Red Check Was a Full DiskSecurity Scanning & CI/CD Gates
- The AI Attack Surface: Four of the Six Layers Nobody ScansAI Security & Supply Chain
- UML: The Boxes Were Right and the Lines Were WrongModelling Notation & Diagrams
- Retrieval-Augmented Generation: Confident, Well Cited, and WrongRetrieval & Search Architecture
- The Scaffolding Comes FirstCI/CD & Delivery Pipelines
- What AI First Actually CostsAI Cost & Procurement
- The Flat LineAI Writing & Detection
- Reasoning Without a ReasonerReasoning & Ontologies
- False GreenAI Code Review
- The Same MAX in the MorningMemory & Identity
- The README Is the PayloadSecurity & Compliance
- Inside MAX's Inside the LoopAgentic AI
- Reading Over the Machine's ShoulderAgentic AI
- A MAX FIELD REPORTAgentic AI
- The 9-Layer Agentic AI Stack: How Modern Organizations Build Run and Secure AI Agents at ScaleInfrastructure
Engage
Building one, or trying to trust one somebody else built?
Architecture review, an evaluation that actually bites, a second opinion on an agentic design, or help getting one through a compliance conversation.
This site: HTTPS only · no third-party trackers · no analytics cookies · security.txt