AI agents
22 articlesTwo vendors, one standard: running a real upgrade with synthesis engineering
A real three-day upgrade with two AI vendors exposed the Built/Deployed Gap: controls and records that passed checks but never reached the running system.
The Decision Packet: batching AI agents' questions without rubber-stamping them
A synthesis engineering pattern for staying in command of AI agents at speed: collect every decision they owe you into one honest, reviewable page.
The migration that was finished until a stranger looked
A synthesis engineering story about Blind-Spot Rotation: how an AI reviewer with fresh context can find gaps the original inventory and checks missed.
“Not found” is not “does not exist”
An AI agent's 'not found' is not 'does not exist.' Five rules for proving absence with bounded reads, positive controls, and scoped negative claims.
Why Synthesis Engineering Exists: Building Across AI Agent Runtimes, Not Around Them
Why synthesis engineering exists: model, runtime, and state portability across Claude Code, OpenAI Codex, Cursor, Hermes Agent, and OpenClaw, without lock-in.
My AI executive assistant writes in its own voice. My AI chief of staff writes in mine.
Ragbot and Ragenie, EA AI and chief-of-staff AI: why my executive assistant agent writes in its own voice while my chief of staff agent writes in my voice.
My design record ruled out agent-to-agent messaging. Six days later I shipped it.
I documented why session-to-session agent messaging would fail, then built it six days later. Why a design record is not a constraint, and what actually is.
What agent memory research taught me about twenty years of notes
Tiered memory for a personal knowledge base: what MemGPT, Zep, and Mem0 teach about structuring two decades of notes so AI agents find the right ten files.
AI tools should copy GitHub's identity model, not your employer's
Enterprise workspaces hand you an account and take it back on departure. GitHub does the opposite. Multi-user AI tooling keeps picking the wrong one of the two.
What should survive when you change AI tools
Project state, methods, authority rules, coordination, and evidence: five things open agent infrastructure should preserve across AI tools and clients.
A Green Config Is Not a Working Agent: Five Planes of Agent Conformance
A green config is not a working agent. Five conformance planes for AI agent infrastructure: source, installed state, live behavior, continuity, capability.
I gave my AI agents a chief-of-staff job description. The hard part was the org chart, not the code.
AI agents as chief of staff: how I encoded absence coordination, notification tiers, send gates, and silent-failure validators into open-source config.
Ask Your AI to Refute You, Not to Review You
AI review defaults to confirmation. A refutation protocol: hostile lenses, uncertainty charged against the work, findings that must survive their own panel.
Honest labels for AI-assisted correspondence
Honest labels for AI-assisted correspondence: three lanes that tell recipients whose words they're reading, from executive staff practice to Ragbot and Ragenie.
The AI agent that deleted 474 of my files was right. I checked anyway.
An AI agent deleted 474 files and its report was accurate. Verifying an agent's summary, not just its work, is the discipline that catches the divergence.
A Warning That Doesn't Block Is a Bug
A pre-commit guard printed its own failure and passed the commit anyway. The fix for the class: fail closed, carry a doctor, spend false alarms like currency.
For AI disclosure guards, the secret isn't the name. It's the approval.
Name blacklists over-block your published biography and under-block real leaks. A disclosure model for AI agents built on approval, provenance, and five tests.
I told a sub-agent to check before committing. Minutes later, I didn't.
Multi-agent AI coding creates failures solo sessions never see: a shared git index, one-shot briefs, report chains. Three incidents and the checks they taught.
I built the conformance validator Google's OKF repo didn't ship
Google's knowledge-catalog repo shipped no standalone OKF validator, so I built a checker for the format's three hard rules. Open-sourced as synthesis-okf.
Why my AI stack has three runtimes
Conversation, procedure, direct manipulation: why my AI setup runs three interaction modes over one substrate of plain files I own, not one chat box.
Your coding agent should watch its own work run
Your coding agent should watch its own work run: browser automation like Playwright turns 'it should work' into 'I watched it work' with evidence attached.
Synthesis Project Management Now Works Across Claude Code and Codex
Synthesis project management now works across Claude Code and OpenAI Codex by keeping durable project memory in CONTEXT.md, REFERENCE.md, and sessions. Updated