Log
Technical articles and lessons from a decade of building web applications.
· 2 min read
What I’m building with sched.fyi
sched.fyi is my booking platform for service businesses in the Philippines and New Zealand: one booking link, customer confirmations, and SMS reminders.
· 9 min read
The CVE Anthropic Calls A Feature
An 8.6 HIGH command-injection CVE in Claude Code, marked Disputed two weeks later. Anthropic says it's how non-interactive mode is supposed to work. Phoenix Security says env var control is reachable via .env files, CI/CD, and IDE settings, no code execution required. Both are right. Which one applies depends on where you run Claude Code.
· 10 min read
How To Teach Your Agent To Dream
Your Claude Code transcripts pile up in ~/.claude/projects/ and your next session knows nothing about them. The orchestrator dream pipeline mines those transcripts, extracts durable learnings, and injects the best ones into the next session before your first prompt. Four steps to wire it up. Plus the four things this pipeline still can't capture.
· 10 min read
How To Compact Before Claude Code Compacts You
A status line script and a session-handoff skill. Two files, under 200 lines of bash and markdown combined. Copy them into your Claude Code config and you never have to wait for /compact to auto-fire at 95% again.
· 9 min read
Your Agent Doesn't Need Six Retry Layers
Six layers of retry logic inside a worker loop is one valid answer. Zero is another. The decision that matters is where resilience lives, not how many layers you stack. Here's the rule I use, the failure table that enforces it, and the test that catches it when someone on your team puts a retry in the wrong place six months from now.
· 9 min read
Your Agent Doesn't Have A Reasoning Problem
Everyone is reaching for Opus 4.7 when the real problem is that the agent can't find what's already stored. I spent two weeks fixing that in my own system. Here's the nine gaps I fixed, why each one mattered, and why most teams building agents will end up doing the same work.
· 8 min read
Your Scheduled Jobs Are Lying To You
I asked my agent system for fresh intel and it handed me a five-day-old brief. The scheduled job had been running successfully the whole time. It was writing to a directory nothing else reads. This is a how-to on catching the failure mode where exit 0 tells you nothing.
· 9 min read
Prompts don't give you voice. Context does.
I spent a week building out a voice system for my AI agents. More rules didn't fix the problem. What fixed it was giving the model evidence of who I am, not instructions about how I want to sound.
· 8 min read
Your Agent Is Learning While You Sleep
250 transcripts. 2,274 learnings. Four structural blind spots. Here's how the dream cycle works — and where it quietly fails.
· 7 min read
Most Of My AI Workers Vanish. One Doesn't.
Most AI agents in my orchestrator are disposable by design. One is named alpha. Here's why alpha exists, how it differs, and what I think it could become.
· 20 min read
How to Build a Harness Anthropic Won't Kill
OpenClaw's cheap subscription run ended April 4. OpenCode had preempted the kill with a nineteen-minute PR titled 'anthropic legal requests.' The stated reason is cache economics. Here are five rules for building a harness that survives.
· 13 min read
I Built Perplexity Computer. You Can't See It.
Perplexity shipped their agentic computer product. Their sidebar is my feature list. What's missing isn't code — it's the label on the box.
· 11 min read
Agent Memory Is Mostly Forgetting
I said memory was a pattern, not a product. I stand by that. But the last four weeks were spent building the fence — and almost every line of it was a way to forget.
· 11 min read
Self-Improving AI Is Boring. That's the Point.
Every demo of 'self-improving AI' looks like magic. Mine is 4,038 lines of Go, three SQLite databases, and a lot of thresholds.
· 11 min read
31,000 Stars and No Orchestrator
I ran five parallel analysis missions against hermes-agent. What I found wasn't a competitor — it was the other half of the stack.
· 7 min read
Your Agent's Persona Is Making It Worse
I spent six weeks building an orchestrator that treats personas as constraint sets, not job titles. Here's what changed and why.
· 4 min read
The newsletter you signed up for doesn't exist anymore
It split in two. Same subscription, two tracks — one for what I'm building, one for what I think about the industry. No action required.
· 9 min read
Stop Running Your Agents
A Reddit post about a 3-agent Claude team hit 427 upvotes and 226 stars in 3 days. The pattern is right. The gap it leaves is the interesting part.
· 17 min read
From 7,415 Lines of Bash to a Personal Intelligence OS
How Via evolved across four generations in 87 days — from Redis polling loops to a 40-skill orchestration engine with 55,000 lines of Go.
· 6 min read
How I Automated My Entire Substack Workflow
From drafting articles to scheduling notes to engaging with readers. One system runs all of it.
· 9 min read
Multi-Agent Orchestration Is Solving the Wrong Problem
The HN wave of 'agents as teammates' assumes the team metaphor improves output. The production data suggests it improves comprehension instead.
· 9 min read
Most Agent Failures Aren't Reasoning Failures
97% of a subprocess output vanished between runs, and it took two weeks to realize every bad agent decision traced back to context, not cognition.
· 9 min read
Agent Frameworks Are Solving the Wrong Problem
Frameworks optimize for orchestration plumbing when the real bottleneck is context management.
· 9 min read
541 Hours of Claude Code in 20 Days
82% success rate sounds great until you examine the 18% — entire sessions lost to context limits, wrong approaches, and one mission that generated 1.7GB of garbage output.
· 8 min read
Isolation as Configuration
Claude Code's worktree support isn't a workflow convenience — it's the first declarative isolation primitive in multi-agent coding tooling, and it changes what parallel execution actually means.
· 8 min read
Your CLAUDE.md Is Making Things Worse
A newly published study found that LLM-generated context files hurt task completion and raise costs by 20%. The harder question is when they still make sense.
· 9 min read
Why the Decomposer Always Picks Backend-Engineer
83 of 275 phases in Via's mission history were assigned to backend-engineer — 30.2% of all decomposed work. The second-place persona got 36. Here's why that gap exists and what it costs.
· 7 min read
The Platform Came Last
I didn't design Via as a platform and then write plugins for it. I built nine CLI tools that each solved one problem, and the platform was what emerged when they needed to talk to each other.
· 8 min read
How the Orchestrator Actually Works: 7 Packages, 4,570 Lines, Zero Magic
I typed orchestrator run and watched three agents spin up in parallel, each writing to its own directory, none aware the others existed. This is the 4,570-line Go system that makes that work.
· 10 min read
MCPs Are Dead. CLIs Won.
A Hacker News post declared the protocol wars over. I've been building on that thesis for months — 175 missions, 9 plugins, zero MCPs. Here's what I actually learned.
· 8 min read
My AI Agent's Memory System Was 60% Dead
I audited Via's five memory layers against Google's memory taxonomy. Three hooks silently failed. MEMORY.md was structurally bypassed. Only one layer — the learnings database — was actually persisting knowledge across sessions.
· 11 min read
The Vampiric Effect
I shipped an 847-line PR on a Wednesday night and couldn't recall a single architectural decision the next morning. The output was mine. The understanding wasn't.
· 12 min read
The #1 Thing My AI Agents Learned Wasn't Code
I built a system that captures what AI agents learn across 175 missions. The most-used learning — retrieved 247 times — isn't a code pattern. It's a publishing decision.
· 6 min read
Three Tools I Built Because Someone on the Team Was Stuck
A CS team that couldn't reproduce bugs, QAs blocked by stale cache, and translations that wouldn't update. Three frustrations, three tools.
· 8 min read
What Nobody Tells You About Migrating a Live Trading Platform
What it's actually like to modernize a live trading platform when you can't break anything and you're the only frontend dev.
· 7 min read
Eight Months Old and Already Load-Bearing
I built my thesis on Docker 0.8. The tool barely existed. The instinct I built using it has never stopped.
· 7 min read
The Patterns From My First Job
A remote hire in the Philippines, inheriting an existing product. Four inconsistent APIs on a client project I wasn't supposed to be building. Docker before anyone in the room believed in it. The patterns that formed before I had words for them.