April 1, 2026
The Path to Autonomous Agents Was Mapped Decades Ago. Nobody Noticed.
→ anatoly.com
A practical methodology for building autonomous AI agents — not by making models smarter in isolation, but by observing how humans guide AI in real conversations, extracting the control patterns, and replacing them one at a time with programmatic equivalents. Draws on decades of practice from aviation, call centers, and education.
- The Control Protocol You Didn't Know You Had
- What Toyota Knew That We Keep Missing
- The Tools Are Here. The Protocol Is New.
- A Switch Would Kill. They Built a Dial.
- The Seat Next to the Expert
- The Blackboard Had It First
- Maybe the Human Was Never Supposed to Leave
- Simple Workflows Bend to AI Easily. Enterprise Operations Don't.
- Hunt the Hunter
March 24, 2026
AI Reads Every Word You Say. It Still Gets You Wrong.
→ anatoly.com
Why even the most carefully worded prompts and rules fail to capture intent — and what to do about it. This post examines the fundamental gap between specifying instructions and conveying meaning, arguing that better structure (not more rules) is the path forward.
- The Specification Problem
- AI Doesn't Have Your Common Sense
- The Confidence Problem
- You Can't Even Watch It Fail
- The Rules Trap
- Intent Over Instruction
- The Harness: Better Structure, Not More Rules
- The Framework: Beyond Prompts
March 18, 2026
One Million Lines of Code. Zero Keystrokes. Welcome to Harness Engineering.
→ anatoly.com
An in-depth look at the emerging discipline of harness engineering — the practice of wrapping AI models in layered structures that control prompt context, architectural constraints, entropy, and verification loops to produce reliable, production-grade software at scale.
- The Drone Operations Center
- What a Harness Actually Is
- The Three Nested Layers: Prompt, Context, Harness
- Context Engineering Layer
- Architectural Constraints
- Entropy Management
- Verification and Feedback Loops
- Security
- Frameworks and Tools
March 11, 2026
One Sentence Can Hijack Your AI. Here's How to Stop It.
→ anatoly.com
A comprehensive guide to AI security in enterprise settings, covering the attack surface of AI agents, the top three attack vectors (direct injection, indirect injection, agent-to-agent propagation), and six practical defense techniques drawn from military intelligence and zero-trust principles.
- The Attack Surface
- The Top Three Attack Vectors
- Trusted Data as the Real Threat
- Zero-Trust Security
- Technique 1: Compartmentalization
- Technique 2: Source Verification
- Technique 3: The DMZ Architecture
- Technique 4: Human-in-the-Loop
- Technique 5: Observability and Audit Trails
- Technique 6: Rate Limiting and Anomaly Detection
March 5, 2026
100% AI Code at Anthropic. 19% Slower Everywhere Else. Why?
→ anatoly.com
Explains the dramatic gap between AI-native labs reporting near-total AI code generation and the measured slowdown experienced by most engineering teams. The difference comes down to architectural complexity — and the post argues that every engineer must now develop real architectural thinking to use AI effectively.
- Greenfield vs. Maintenance
- Enterprise Complexity
- Enforcement vs. Architecture
- The Cognitive Juggling Limit
- Why Every Engineer Needs Architectural Thinking
- The Limits of AI Skills
- Keep Coding
February 26, 2026
Two Roads for AI in Software Engineering — and Neither Is What You Think
→ anatoly.com
Identifies two emerging use cases for AI in software development: bottom-up (AI as a developer productivity tool) and top-down (AI building entire systems from a prompt). Examines trust, risk compensation, compartmentalization, and domain-specific languages as key considerations.
- Bottom-Up: AI as Productivity Tool
- Trust and Risk Compensation
- Compartmentalization in Architecture
- Top-Down: AI Building Entire Systems
- Domain-Specific Languages
- Formalizing the Handoff Between Phases
February 26, 2026
Beyond Chatbots: The Case for AI-First Software Architecture
→ anatoly.com
Makes the case for purpose-built AI-first architectures by walking through the challenges of integrating AI capabilities — phone calls, SMS, decision-making — into traditional enterprise systems. Covers long-running activities, human-in-the-middle workflows, security, and the shift from deterministic to probabilistic software.
- Autonomous AI Phone Calls
- Long-Running Activities
- Human-in-the-Middle
- Security and Zero-Trust
- AI Decision-Making in Applications
- Deterministic vs. Probabilistic Architecture
February 26, 2026
The Bedridden Genius: A Mental Model for What AI Can Actually Do
→ anatoly.com
Develops a practical mental model for understanding AI capabilities and limitations: think of the LLM as a statistically average person with tools (aides) and a persistent overconfidence problem. Covers what AI can and cannot do for both business operators and software engineers.
- LLMs as Statistical Text Prediction
- Tools as "Aides"
- The Dunning-Kruger Effect in AI
- AI as a Junior Engineer
- What AI Can Do for Business
- What AI Cannot Do
- What AI Can Do for Engineering