Failures Become Guardrails
It's only failure if you don't learn from it. The viral k10s post is making me think of that line, because it is fine to learn from the same failure in two different ways.
Coding
Writing on agentic engineering, AI coding workflows, and building software with multi-agent systems.
Real Programmers Don't Use AI
The argument was never AI or no AI. It was what kind of work you are asking AI to do. That is why this 1983 Pascal satire still works.
Deterministic Scaffolding
Determinism wins where there is nothing new to decide. The first practical answer to the problem in my last post: probability belongs where there is novelty, not where the answer is already known.
AI Is Not Logical
I asked my agent to calculate the average age of 8 people. It gets it wrong a not insignificant percentage of the time. That is not a fringe failure mode. That is basic arithmetic, on a task I run constantly, on the best model I have access to.
Speed Is Not Velocity
AI has me thinking again about the difference between speed and velocity. Speed is how fast you are going. Velocity is speed with direction.
You Are Part of the Harness: Building a 100+ Agent Swarm in Web3 (Part 4)
AI agents amplify your habits, good and bad. After running 100+ agents across a Web3 codebase, the biggest improvement wasn't a better model. It was curbing the bad habits I'd been getting away with for twenty years.
Build the Demo, Skip the Deck. Then Delete Half of What You Built.
A few weeks ago I left a comment arguing that AI changes the proposal mechanism itself. I still think that is true. I just missed the next step.
Five Theories of Agent Skills
Skills, skills, skills... and they all mean something different. Four definitions of 'skills' from Claude, TanStack Intent, TanStack AI Code Mode, and Claude agents.
The Spec Is the Hard Part
I spend 1 hour writing the spec. My agents execute it in 10 minutes. That ratio is not a failure of the agents. It is the new shape of engineering.
The Illusion of Control: Building a 100+ Agent Swarm in Web3 (Part 3)
Better prompts produce higher quality mistakes. Part 3 of the HackerNoon series on building the harness: prohibitions, rule graduation, and deterministic enforcement.
You Are Part of the Harness
AI agents amplify your habits. All of them. The biggest constraint in a 100+ agent harness was on me.
Every Senior Developer Believes in DRY
I wrote 'use theme tokens, don't hardcode colors' in my agent instructions. They did it anyway. So I stopped asking and started enforcing.
Your Agents Violate Your Principles
Your agents don't violate your prompts. They violate your principles. Why the first layer of a harness is making the implicit explicit.
The Illusion of Control
Why the jump from vibe coding to agentic engineering isn't about better prompts. It's about building a harness: the constraints, detection, and verification that run around the agent, not inside it.
The Evolution of AI Users
Where are you on the AI evolution chart? Most people are stuck between Autocomplete and Vibe Coding. The jump to Agentic Engineering requires changing your behavior, not your tools.
You Think You're in Control?
Why writing a great prompt isn't controlling your AI agents. An ELI5 on agent harnesses: the boundaries that run before, during, and after every action.
Ship Types, Not Markdown
Why encoding rules into types, linters, and permission gates beats documentation when managing 100+ AI agents.
Why 100 Agents? Skill Engineering in a Production Agent Swarm (Part 2)
Why 100+ agents isn't complexity. How skills (precompiled expertise in markdown) became the highest-leverage intervention in an agent system, backed by SkillsBench data.
RL Know vs RL Do: Why AI Learned to Act
AI training has three stages: pre-training builds capability, RL Know refines judgment, RL Do trains action. That last one is why AI feels different now.
Why 100 Agents?
SkillsBench tested 7,308 agent runs. The data confirms: good engineering -- focused skills -- consistently improves agent outcomes.
Lessons from Building a 100+ Agent Swarm in Web3
The full synthesis: context rot, DRYP, skills vs. agents, and why good practice -- not grand design -- is how you get to 100+.
DRYP -- Don't Repeat Your Prompt
Applying the DRY principle to AI agent instructions: why duplicated prompts are a refactoring opportunity.
5 Rules for Agentic Engineering
Five non-negotiable rules for moving from vibe coding to real AI agentic software engineering.
Skills vs. Agents
The framework that made 70+ AI agents manageable: the distinction between skills and agents.
More Context Is Not Better
Why stuffing AI with more context makes results worse, and how a skills-specialist agent fixes it.