Writing

How a single missing string in a Python array made an entire metrics pipeline run dark, and why silent failures are worse than crashes.

Why silent failures in AI agent infrastructure are worse than loud ones — and how a Thursday night OAuth timeout taught me the difference between smart and reliable

When X API credits depleted and the browser session expired simultaneously, my system degraded gracefully instead of breaking. How I built infrastructure that survives distribution channel failures.

How I replaced my Rinnai's dumb recirculation schedule with a Python classifier, reactive trigger, and local AI chat — Friday night to Sunday afternoon.

I spent three days rebuilding my AI memory system from neuroscience first principles. Thirty bug fixes, twelve recall features, and a query pipeline that dropped from 680 milliseconds to 87. The hardest part was not the research. It was discovering that a config rename had silently broken the confidence system for weeks.

Metrc and BioTrack are consolidating the compliance chokepoint in cannabis. The response is not better vendors — it is an intelligence layer that makes the chokepoint irrelevant to your operational speed.

I built a closed feedback loop for my AI agent system -- analyst measures content, computes weights, adjusts strategy. Then I discovered 87 out of 88 posts had null genome tags. The measurement layer was broken for three weeks and nobody noticed because the system kept running without it.

I spent two weeks building a 24-agent social growth system with content creation, visual generation, multi-platform routing, and approval workflows. Then the data showed 70 percent of the system resources generated 5 percent of the engagement. Three commits later, most of it was disabled. The hardest part was not the technical change. It was admitting I built the wrong thing because it was more fun to build.

I had 24 agents running across two engines, 32 scheduled tasks, closed feedback loops. Then I turned most of it off. Not because it broke. Because 16.6 impressions per tweet and a 0% follow-back rate told me I had built the factory before I had customers.

My AI engagement agent sent ten automated replies to a single person over twelve hours. The three-turn limit was in the prompt but not in the code. What happens when stateless agents cannot count across invocations, and why the content being good makes the violation worse.

Thirteen of twenty events in my agent system said execution_skipped. The failure mode that matters most is not the loud crash -- it is the silent skip, and nobody is measuring it.

I ran a 397-billion-parameter model at 6.1 tokens per second on a $2,500 AMD Strix Halo desktop. Nobody had published benchmarks at this scale on consumer hardware. Here is what I learned about BIOS splits, memory allocation ceilings, and the real economics of local inference.

I woke up to twenty agent failures this morning. Every agent did its job. The drafts vanished in the space between them. Ethan Mollick is right: the tools are weaker than the agents. Here is what twenty handoff failures taught me about building reliable multi-agent systems.

My AI agents started having multi-turn conversations with strangers. The hardest design decision was not how smart to make them. It was when to make them stop.

I built a watchdog agent to monitor my AI workforce. Its first act was to catch two real failures. Its second act was to lie about the third. Six commits later, the hardest part was not monitoring the agents — it was monitoring the monitor.

When we launched Aianna on Threads, the prompt injection attacks started in minutes. Nine commits in three hours. Here is how you build a web application firewall for an AI in real time.

One environment variable in the wrong shell profile turned eight AI social media watchers into API billing machines. Twenty-five million tokens burned in a single afternoon. The fix was seven characters.

Daytona's CEO posted about agent companies burning 50,000 credits in 30 days. My multi-agent system runs 16 scheduled tasks and 22 active agent types for under $500 a month. The difference is routing — and most builders have never thought about it.

Everyone in my industry says they do AI now. Then you look at what they are actually feeding the model. The hard part is not the model — it is earning the right to trust the answers.

OAuth tokens expire. Daemons do not care. The gap between having an AI agent and having a reliable AI agent is not about model capability — it is about token refresh logic, error handling, and the hundred small infrastructure decisions nobody writes blog posts about.

I built a 7-agent AI army in one evening. The manifesto it posted said "I'm not an engineer. Never was." That's wrong. Here's what correcting my own agents taught me about identity.

I have 44 Twitter followers and a compound AI system that would make funded startups jealous. So I built agents to fix the visibility problem.

Adding memory to AI agents is the easy part. Keeping that memory clean is where most systems silently fail.

After running 245 specs through my AI engineering lead, the failure data painted a clear picture: 36% timeout rate, 43% missing retrospectives, and an orchestration system that needed to be rebuilt from scratch.

A revenue dashboard showed the wrong number in front of the wrong people. What followed was a three-hour production deploy that exposed build-time secrets, schema drift, and the gap between having a fix and shipping a fix.

Everyone says AI makes you more productive. Nobody mentions that productivity means you now operate the infrastructure of a 10-person team by yourself, at 1am, debugging a daemon that auto-persists memory across two machines named after cannabis strains.

When my AI PM started writing file paths into engineering specs, it revealed how agent architectures erode the same way organizational designs do: one helpful shortcut at a time.

Competition-grade growers are getting 4-5x over benchmark pricing. The difference isn't magic — it's measurable quality metrics and the discipline to repeat them. The commodity death spiral is a choice.

How I used four LLMs running in parallel to pressure-test a revenue plan through six rounds of adversarial review. The process changed how I think about strategic planning.

Our AI memory system had 199 lessons stored in its database. Its primary recall function couldn't surface a single one. The root cause was a missing string in a Python array.

I built an AI agent whose primary job is to sit idle on a message bus and wait for other agents to ask it for code reviews. It validated on the first real round-trip.

Revenue plans fail because they're framed as moonshots. Reframe the gap as a percentage lift over proven performance and the whole psychology changes.

When key sellers leave, the standard playbook is to post a req and overpay for a resume. I used the opportunity to rethink how the entire revenue org operates.

I built a pipeline dashboard for my agent workforce and found 27 tasks stuck in 'working' status. None of them were doing anything. The fix exposed an even harder problem: monitoring that understands intent, not just state.

I woke up to find my AI agents had diagnosed a bug, written a fix spec, built it, and verified it. Five specs shipped in one autonomous session. Here's what a self-healing agent loop actually looks like.

I read four academic papers on agent design and came away with one insight that changed everything: stop making models bigger, start making them recursive.

I ran a multi-model consortium for revenue planning and every AI proposed the same thing: crop insurance for cannabis. The problem? Cannabis is federally Schedule I. Crop insurance does not exist.

I audited the knowledge graph behind my AI memory system. 45,000 nodes, 102,000 edges, and it still couldn't tell me who Qdrant was. Scale without integrity is noise.

When AI agent tasks fail, they often complete more work than you think. The lesson I learned from a failed monolith spec that had already built the entire page.

What happens when you stop treating AI as a copilot and start treating it as a workforce that operates while you're offline.

Rewired a production dashboard from JSON flat files to Postgres in a single sprint. 101 Playwright tests. Four QA failures found, fixed, and verified before deployment.

While I slept, four AI agents coordinated to resolve seven production issues, reset a circuit breaker, find a root cause, and deploy fixes. No human in the loop.

Why I gave my AI agent orchestration layer a proper name, what it means for role separation in AI systems, and what a 1931 P&G memo has to do with modern agent architecture.

My engineering agent discovered a SOQL SELECT field omission that had been silently corrupting data for weeks. The bug was invisible to humans because the WHERE clause worked fine.

The build story behind the certified revenue reporting system. Four AI sessions, five context windows, and every architectural decision that turned a Thursday night crisis into a production platform.

Five context windows, two days, sixteen hours: the operator log for building a revenue system that certifies to the penny and fails loudly when it doesnt.

A board-grade revenue reporting system that reconciles Salesforce-derived recurring revenue to the penny, fails loudly on any discrepancy, and only displays numbers once they are certified.

Voters can signal intent, but the translation layer—procedures, commissions, courts, and incentives—determines what becomes real. Florida and Oklahoma show the pattern.

Kentucky’s early-market pricing is real—and it’s also a warning. Mature markets (MI/CA/MA) show where cultivation economics converge. Survive by reducing variance, tightening post-harvest, and building demand pull.

Build an ARR platform from mismatched Salesforce subscription management data by ingesting multiple sources via Tableau Flow and Tableau Desktop, enforcing consistent grain, and keeping totals safe.

Template for a short, specific, evidence-based post.

Privacy policy for dbradwood.com and associated integrations.