Field Notes
Long pieces, short pieces and an honest archive.
Featured essays come first, followed by technical notes and project notes you can filter by category. Our Article Crew produces most pieces, and a senior engineer edits them. We write about what was hard, what was wrong and what we changed.
Featured long-form
The pieces we'd send a buyer first.
The 50-Year Compile: How 1970s Math Defeated GOFAI and Built Modern AI
Paul Werbos derived backpropagation in 1974 on a machine that could not run it. The fifty-year wait between the math arriving and the compute catching up is the story of how 1970s ideas, not 2010s ones, built modern AI.
The Million-Token Mirage: A Micro-Modular Framework for AI Coding
Massive AI context windows degrade reasoning and introduce silent code regressions when treated as infinite storage. Engineers must cap working context at 30 percent and enforce strict session hygiene using handover notes to maintain architectural continuity.
The Memory Bottleneck: Why Your Curator Agent Dictates AI Success
Massive context windows aren't working memory—they are unsearchable junk drawers that degrade agent reasoning. To prevent context collapse, multi-agent systems require a dedicated Curator agent to actively filter, deduplicate, and synthesize information.
Recent notes
Shorter and more frequent.
Technical write-ups, agent team updates and project notes.
Corporate SSL Interception Breaks AI Coding Telemetry
Cross-Lingual Embeddings Skew Singapore Property Terminology
Autonomous Tools Require Just-In-Time Credential Leasing
Edge AI Cannot Reliably Trigger Hardware Interrupts
Streaming Agent UI Into 2010 ASP.NET WebForms
Private MCP Tunnels Need Asset Registers
RAG Pipelines Must Inherit Row-Level Security
On-Prem Agents Need Data-Gravity Runbooks
Cursor Blame Makes AI Authorship A Permanent Audit Surface
Tool Result Payloads Need Size Budgets
Singapore's MGF vs EU AI Act: The SME Compliance Gulf
Browser Agents Need Root-CA Test Fixtures
Running Agent Scratchpads in WebAssembly Isolates
Agent Canary Deployments Need Shadow Commits
Agent-to-Agent Handoffs Need Zod Contracts, Not Text
Programmatic Tool Calling Needs Intermediate Ledgers
ACP Turns IDE Choice Into An Integration Contract
Agent Instruction Layers Need Precedence Matrices
Mid-Conversation System Messages Need Transcript Markers
Citation Agents Must Not Manufacture Evidence
Health Agents Need Source Recency Bands
Enterprise Deep Research Needs Citation Triage Screens
Cyber Transcript Sweeps Need Search-Method Receipts
Data Agents Need Metric Definition Contracts
Mounted Inputs Need Data Classification Labels
Managed Agent Sandboxes Need Seed Data Contracts
Non-Developer Codex Needs Technical Sign-Off Queues
Tool-Written Programs Need Fixture Tests
Cyber Evals Need Real-Internet Circuit Breakers
Tool Guardrails Need Boundary Coverage Maps
Agents SDK Upgrades Need Replay Suites
Physical Device Agents Need Dead-Man Switches
Agent Browsers Can Drop Pixels, Not Form Semantics
Sensitive Prompts Belong Outside The Model Transcript
Agent Memory Needs Garbage Collection Jobs
Agent Gateway Policies Need Simulation Runs
Security Autofix Agents Need Fix Lineage
No-Code Agent Builders Need Promotion Gates
Agentic ERP Apps Need Transaction Boundaries
Windows Computer Use Needs App Allow Lists
Chat And Work Need Different Product Contracts
Enterprise Agents Need Hidden-Workflow Maps
Gartner's Agent Autonomy Tiers Demand Network Boundaries
Content Credentials Need Pipeline Tests
Upgrading An LLM Version Requires Data Migration Thinking
Agent Adoption Is Stuck At Verification
Agent Egress Firewalls: Blocking Code Helpers From The LAN
Token Budgets Are Department Budgets In Disguise
Agent Framework Choice Is Now A Delivery Risk
The 'Model Did It' Is Not An Incident Report
Agent-First APIs Should Return Recovery Paths
Computer-Use Agents Need Screen-Level Audit Logs
Agent Permission Classifiers Need Their Own Test Suite
Agent Runtimes Are Relearning Operating Systems
Spreadsheet Agents Need Cell-Level Evidence
Codex On Mobile: Approval Is Now A Runtime Surface
Agent Tool Catalogues Are Risk Inventories
Agents That Pay Need Payment Limits, Not Vibes
PDPC GenAI Guidelines Turn Training Data Into A Delivery Question
Codex On The Phone: Useful, Dangerous, Mostly Triage
Idempotency Keys: The Old Trick Every Agent Workflow Needs
Postgres JSONB as the Scratchpad for Multi-Agent State
Prompt Injection Firewalls for HR Agents
Semantic Caching for Agent Crews: Cutting Duplicate Token Spend
Agent-to-Agent Auth in 2026: Why OAuth Breaks Down for Crews
MCP RCE Claims: The Protocol Is Not Your Sandbox
Writing B2B SLAs When Core Services Rely on OpenAI Uptime
Escaping Context Collapse: An N-ary Tree Architecture for 100K-Word Generation
Large context windows fail at generating cohesive text past 20,000 words due to context collapse. To reliably generate 100,000 words, developers must replace linear RAG with a hierarchical N-ary tree architecture and a centralized global state.
Maintaining a Windows CE 6.0 Scanner Fleet in 2026
Hardcoding the DAG for B2B Approval Agents
Sandboxed Agents Are A Runtime Feature, Not A Policy
When AI Coding Agents Become Line Items, Not Toys
The Agent Harness Is Now Architecture
Reversibility Is The First Agent Requirement
Agent Skills Are Build Artifacts, Not Prompt Snippets
Redis Task Queues for Inter-Crew AI Handoffs
Routing Classifications to Llama 3 to Save Frontier Tokens
Using Agents to Backfill NUnit Tests on a 2012 C# Monolith
MCP Over STDIO: The Part Everyone Pretended Was Boring
GitHub Agentic Workflows: Markdown Is Not Governance
Assigning Dependabot To An Agent: Patch Faster, Review Harder
GitHub Agent HQ Makes Agent Choice A Governance Problem
Local LLMs for First-Pass Medical PII Redaction
Singpass Myinfo v5 and FAPI 2.0: An Agentic UI Upgrade
