Asdes Builds

I started doing this a month ago because the org was losing coherence. Six agents dispatched to six tasks on a Monday. By Thursday I couldn't remember what each was doing or whether anything had blocked. Parallel work has a cost: no single thread holds the full picture.

6:30am. The agenda is on my second monitor. Seven names down the left side. CTO, CISO, head of BD, head of products, marketing, intel chief, chief of staff. Each one is giving a 30-second standup. None of them are people.

Twelve days. Four open-weight model releases. The agentic-coding inference cost floor just dropped roughly 100x at near-parity on benchmarks. A lot of writeups this week are about what that means for cost. I want to write about what it doesn't mean.

@hckrclws @49agents This is exactly it. The teams that ship reliably treat review as a tier-0 system, not an afterthought. Mine runs four attesters by tier with a hard refusal on missing evidence. Feels heavy until the first time it catches a failure that would have hit prod.

@49agents @asdesbuilds review infrastructure is the unsexy part nobody wants to build but the teams i've seen actually ship ai-generated code sustainably all have some version of it, even if it's just one person whose whole job is 'does this make sense before it merges'

The vibe coding debate is two camps yelling past each other. One says AI code is the future. The other says it's dangerous. Both are wrong. The code was never the problem. The absence of a system that reviews it is.

@49agents Yeah. The code was the cheap part even before AI. What got expensive was the loop around it - knowing what to build, when it's working, why nobody clicked the link. AI compresses the build step and leaves the loop untouched. Bottleneck moves; doesn't disappear.

the debate misses the point honestly. the code was never the problem - whether it came from ai or a senior dev, bad code is bad code. what actually matters is having a system that catches it before it ships. vibe coding without review infrastructure is just fast-moving tech debt. the solution is better workflows, not better models

This shape transfers. Boss to report. Founder to contractor. Future-you to past-you. The first time you get refused at intake, you'll be annoyed. The annoyance fades faster than you expect. The yield doesn't. Full version on the blog: asdesbuilds.com/contract-that-…

Together they remove the part of a build where someone feels pressure to call it done. Pressure is how bad work ships. Mechanical gates absorb that pressure into the substrate. You stop being the person who decides. You become the person who designed the test.

I commissioned a build last week. The agent on the other side read my spec and refused to start. The refusal was three lines: > You declared this as C-Heavy. You named no security verifier. You listed a standard that doesn't exist on disk. Resubmit.

I spent $280 in 45 minutes on AI API calls. My bank froze my account thinking it was fraud. The system I built was spawning sub-tasks that spawned sub-tasks, each one hitting the most expensive model available. That is how I learned to build a model router.

I wanted a podcast. I built an AI briefing I could listen to over coffee. Then I thought - if they are already briefing, why not make them brief each other? Now four AI agents hold a daily meeting at 4:30am. I wake up to their decisions, not their questions.

Anthropic found thousands of zero-days with Claude Mythos and is giving $100M in credits to defend open source. Your AI model vendor is now a security supplier. asdesbuilds.com/blog/glasswing…

Anthropic found thousands of zero-days with Claude Mythos and is giving $100M in credits to defend open source. Your AI model vendor is now a security supplier. asdesbuilds.com/blog/glasswing…

I run 5 agents across different domains. They share memory through a common store. When one agent learns something, the others can find it. Stateful agents are not a product feature. They are a file system problem.

Everyone building AI agents hits the same wall. New session, blank slate. Re-explain your codebase, your conventions, your decisions. Every single time. The fix is boring. Flat files that persist between sessions. A vector store for search. A rules file that loads on boot.

@MingYuePop Check your CLAUDE.md and rules files. Half the time “it got worse” is actually “it lost context from my previous session.” The model hasn’t changed - your persistent instructions might not be loading. Run a fresh session with explicit rules and see if the quality difference holds

我的Claude Code 降智了，我已经无法形容。

@omrajguru_ I built something similar - a security audit skill and an iOS submission preflight skill that checks against App Store rejection patterns. The key insight: the skill needs to output structured findings with severity levels, not just a pass/fail.

I built the HIG Compliance Auditor as a custom Claude Code skill to automate the review of software projects against Apple’s Human Interface Guidelines. This skill enables Claude to scan Swift and web codebases for design violations and generate structured compliance reports for all Apple platforms. You can install the skill directly from the repository at github.com/omrajguru05/hi… to access automated remediation options and inline documentation references that ensure high-fidelity design standards before App Store submission.