When One Agent Isn't Enough
What happens when agents negotiate with agents? The math becomes visible, the romance disappears, and the questions left over don't have answers.
Agent Protocols: MCP, A2A, A2UI, and When You Need Which
MCP, A2A, A2UI, commerce protocols: they're not layers, they're orthogonal. Here's who's using what in production and when you need which.
The Hybrid UI (When Chat Meets Cards)
Pure chat fails for complex work. The future isn't replacing UI with chat — it's coherent shells with disposable pixels inside. Here's what that looks like in practice.
Where to Put the Agent
My attempt at a framework for deciding where agentic experiences belong in your product. Not everything needs an agent — here's how I think about finding the candidates that do.
The Three Layers That Won't Change
Software is decoupling into three persistent layers: System of Record, Agentic Layer, and Pixels. This isn't prediction — it's already happening.
The Middle Layer You're Missing
Your enrichment data lives in Clay. Your interaction history lives in HubSpot. Your dialer has a 250-character limit. The middle layer is where you reconcile all of it.
Requirements Before Implementation: The Research Flow
TDD works for humans because humans carry domain knowledge. Agents don't. The research flow makes requirements explicit before implementation begins.
Parallel Exploration Without Context Pollution: Sub-Agents
Investigation fills your working context with stuff you don't need afterward. Sub-agents keep it separate — send inputs, receive outputs, don't carry intermediate state.
Work That Survives the Night: Beads for Multi-Session Tasks
Plan mode gives you a document. Beads gives you a graph of work with dependencies that the agent understands. Tasks small enough to survive context loss, blocking relationships that prevent premature work.
Let the Agent Check Its Work: Verification Patterns That Actually Work
If a task is verifiable, it's optimizable. Quality gates that reject bad code, visual testing so the agent can see what it built, and the feedback loop tying everything together.
Claude Code Plugins: From Prompts to Packaged Expertise
A prompt is a one-time instruction. A plugin packages expertise into something that persists — commands you invoke, skills that load on demand, hooks that run automatically.
Claude Code MCP Setup: Teach Your Agent to Search Before It Invents
Adding an MCP server to Claude Code takes one command. Changing how your agent approaches problems takes intention. Here's how to enforce doc-first behavior.
Stop Using Claude Code Like a Copilot
Most developers use maybe 10% of what Claude Code can do. Five pillars unlock the rest: context, skills, verification, persistence, and scaling.
AI for Call Preparation, Not Calling
Technical sales requires preparation, not automation. We spend 10x more preparing for a call than making it. AI handles research. The call itself? That's human.
Clay Can't Be Your Source of Truth
Rows don't know about each other. That single constraint means Clay can't track prospects over time, merge records, or maintain history. Here's why — and what to do about it.
Your Portfolio as Markdown Files
AI coding agents are the most capable tools we have. Your sales knowledge should be queryable by them. Here's how to set it up in 30 minutes.
Don't Trust the Model, Use a Framework
A practitioner's guide to defending against prompt injection. Structured outputs, trust boundaries, and patterns that actually work.
I Broke a 43K-Star AI Assistant in 90 Minutes
Prompt injection is the SQL injection of 2026. Here's why vibe-coding won't save you.
The Agent Skills, Part 1: The Diagnosis
Why 95% of Claude Code skills fail — and what actually works. Most AI skills are system prompts in a trench coat. Here's the diagnosis.
The Agent Skills, Part 2: The Philosophy
Workflows, Not Personas. How to build skills that encode developer behavior, not library descriptions.
The Agent Skills, Part 3: The Architecture
Managing Entropy. How to organize skills so they don't collapse under their own weight.
The Agent Skills, Part 4: The Playbook
How to build Claude Code custom commands that work. The 5-phase process: archaeology, stress test, codification, compilation, and QA audit.