Joey Kudish 🇺🇦🏳️‍🌈

jkudish.comHuman in the Loop NewsletterSubscribe to Human in the Loop for practical insights on AI-augmented coding, productivity hacks, and how to use AI as your development partner, not replacement.

24 Nov 2025

Tomorrow's newsletter: The 80/20 Rule Didn't Go Anywhere AI didn't eliminate the hard work. It just revealed where it actually lives. The easy stuff was never the real work - we were just busy doing it. Now AI compresses it, and we hit the hard parts faster. jkudish.com/newsletter?u...

jkudish.com#014 - Your Failed Experiments Have Expired - Human in the Loop NewsletterIssue #014: Your Failed Experiments Have Expired

9 July

Wrote about my Fable orchestration experience jkudish.com/newsletter/014

sessions in my agent runtime are first-class objects. auto-named, resumable with full context via rolling summary, cross-device (desktop to CLI).

the feature i'm most excited about: session forking. try a different approach without losing your current exploration. plus an @ mention system for injecting context from sessions, tasks, files, notes, people, projects.

multi-agent collaboration patterns i'm building as first-class primitives: fan-out, pipeline, debate, orchestrator-worker.

i use most of these today across a hack of claude code and codex sessions. the difference is doing them manually vs having the runtime handle coordination.

no compaction in my agent runtime. instead of summarizing old messages and throwing them away, i use a sliding window plus a rolling summary. recent messages in full fidelity, older ones replaced by a continuously updated structured summary.

five memory layers: active context, working memory, session memory, semantic memory, and document memory. the goal is an agent that remembers what matters without losing what just happened.

the runtime i'm building is model-agnostic. every agent definition declares a preferred model and a fallback chain. if claude hits a rate limit, it falls back to gpt 5.4 or glm 5.1.

you configure how each agent handles fallbacks: silent for background work, pause-and-notify for interactive sessions, full stop for critical tasks. the best way to build with AI is using the best tool or model for the job.

experimenting with building my own agent runtime. i've customized the heck out of claude code, but it's locked to one model family and missing features i want. figured the best way to learn what i need is to build it myself.

exploring behavioral profiles in my agent runtime. instead of "code mode" vs "chat mode," each profile defines tool access, approval gates, and memory behavior. the UI adapts; what's exposed changes based on the profile.

the profiles: explore, diagnose, design, build, review, devops, operate, converse. each one is a different posture. explore reads but doesn't change anything. build executes a plan. operate runs the business. converse is just thinking together.

on a typical day, or even within the same workstream, i move between most of these. the goal is to make transitions fluid, not something you manually switch. the right tools and UI surface automatically.

13 Apr

librarium is a cli tool i built and maintain that fans out research queries across different providers. I just released v0.1.3 with a handful of new providers including firecrawl github.com/jkudish/lib...

13 Apr

when you fan out a research query to 18 different sources, the consensus answer is the least interesting thing you get back. the disagreements are the signal. that's when the actual thinking starts.

8 Apr

claude code's built-in context tools are good. really good. but they're per-session, per-project. they don't compound across your whole practice. tomorrow's newsletter is about the platform i'm building on top to fix that. jkudish.com/newsletter

7 Apr

most developers write a CLAUDE.md, import a few skills, and stop there. i didn't stop there. wrote about what i built on top in tomorrow's newsletter. subscribe at jkudish.com/newsletter

6 Apr

i stopped thinking of AI as a coding assistant. it's a full implementation partner now. i describe the problem. it asks questions. i approve the plan. it writes the code, tests, opens the PR. my job is the decisions.

6 Apr

treat your CLAUDE.md as a table of contents, not an encyclopedia. short entry point with pointers to deeper docs. architecture in one file. conventions in another. agents navigate better with a map than a manual.

6 Apr

some teams are still debating whether to adopt AI. early ones are experimenting with ways to get reorganized around it. what i'm seeing as successful patterns: simple tasks go straight to agents. manageable tasks get a few feedback cycles. complex tasks stay with engineers.

4 Apr

what does your AI coding setup look like right now? single agent sessions? parallel worktrees? custom commands? something else entirely? curious how people are structuring their workflows beyond "open chat, type prompt."

4 Apr

the best way to make AI agents reliable isn't better prompts. it's fewer choices. test suites, linters, type systems, architectural constraints. every guardrail you add removes a degree of freedom where the agent could fail. agents don't need more intelligence. they need less room to be wrong.

buildermethods.comVideo Editor & Producer — Builder MethodsBuilder Methods is a site that helps software developers build better software, faster by leveraging AI. Created by Brian Casel.

3 Apr

every friday: run a cleanup agent across my repos. scans for stale docs, dead code, outdated patterns from the week. opens fix-up PRs automatically. agents replicate what exists, including the bad parts. regular garbage collection keeps drift from compounding.

Joey Kudish 🇺🇦🏳️‍🌈 reposted

Brian Casel

@briancasel.com

2 Apr

I'm hiring a Video Editor & Producer to work with me at Builder Methods. Know someone? You? Looking for a creative partner to take this to the next level — storytelling, custom visuals, motion graphics. I value taste and creator chops above tactics. buildermethods.com/jobs/video-e...

implementation cost is going to zero. so what's left? vision. taste. judgment. the cognitive demands didn't shrink. they concentrated.

about 30% of my dev time goes to meta-work now. custom commands, agents, feedback loops, memory systems. not shipping features. building the system that ships features. every piece compounds into the next. and eventually, i want other builders to be able to use it too. more on that soon.

the pragmatic engineer surveyed 900+ engineers. 56% now do 70%+ of their work with AI. claude code went from zero to the #1 AI coding tool in 8 months. this isn't early adoption anymore. also who are the other 44% and what's the 30%? am i weird for being at like 99%?

SaaS isn't dying. but it's no longer the default answer to every business problem. Looks like a growing number of businesses are hiring builders to create internal tools instead of buying another subscription. If I can be of help, reach out!

last week i shipped across 6 projects in ~35 hours. not because i worked harder. because AI handled most of the execution. the leverage is real. but i'm still the bottleneck for every prompt. writing about that tension in this week's newsletter. subscribe: jkudish.com/newsletter

7 Mar

built a /deep-review command: fans out code review to 4 AI models in parallel. if 2+ flag the same issue, severity gets bumped. auto-fixes critical and high findings before the PR exists. built on @aarondfrancis's counselors package.

7 Mar

what's the one thing in your AI coding setup that compounds the most? for me it's context. CLAUDE.md files, auto-memory, project-specific configs. every session leaves behind knowledge that the next session builds on. the AI gets better at my codebase without me doing anything extra.

6 Mar

the more i automate my dev workflow, the more my job looks like product management. define the problem. clarify requirements. approve the plan. review the output. the engineering skills didn't become less valuable. they just moved from typing code to systems design.

6 Mar

claude code PostToolUse hooks are underrated IMO. i added a 17-line bash script that detects gh pr create and automatically starts the code review loop. no manual step between "PR created" and "reviewer comments handled." code.claude.com/docs/en/hooks

do you want AI to cite your brand? own the definition. find the specific terms in your niche that don't have clean, authoritative explanations anywhere. write the best one. structure it so AI can extract it cleanly. I call this the "dictionary definition heist."

three layers to my agentic memory system: 1. raw transcripts. every conversation stored as source of truth. 2. compressed observations. what the agent noticed and extracted. 3. normalized truths. facts with confidence scores and validity windows.

truths have a lifecycle. they can be superseded, contradicted, or archived.

most AI conversations start from zero. you open a new chat and the AI has no idea what you worked on yesterday. I built a memory system so mine doesn't work that way anymore. three layers: raw transcripts, compressed observations, normalized truths with confidence scores.

it actively resolves conflicts and automatically resolves contradictory memories. i'm writing about it in the next issue of my newsletter which I'll send tomorrow

jkudish.com/newsletter

i've been experimenting with different AI memory systems. i wanted something that can store observations, connect entities, and retrieve context across sessions. my first instinct was to reach for a hosted vector database. that's what everyone recommends. 🧵

i did evaluate the alternatives. Neo4j has stale Laravel packages. FalkorDB has no PHP client. SurrealDB would need a full Eloquent driver rewrite.

the real find was Apache AGE; a PostgreSQL extension that adds Cypher graph queries. graph traversal on your existing database. no new moving parts. PostgreSQL can handle a lot more than people give it credit for 😅 will share more about what i'm building soon.

instead i ended up building in Laravel with Eloquent and PostgreSQL. pgvector handles embeddings. regular tables handle the entity graph. Laravel AI SDK handles extraction. everything runs inside the same database my app already uses. no new services. no new infrastructure to monitor.

caught up with the latest development happenings in WordPress over the weekend. some of my clients are heavily invested in the ecosystem and it's important to stay current. the biggest shift: the WordPress Interactivity API (IAPI). 🧵

the other thing that caught my attention: the official WordPress/ai plugin. it adds an Abilities API that makes WordPress discoverable to AI agents. register capabilities, connect to OpenAI/Claude/Gemini, and let AI interact with your site programmatically. this is early but it's a foundation.

between IAPI for interactivity and the Abilities API for AI, WordPress in 2026 is not the WordPress people dismiss. it's evolving fast and the AI integration story is stronger than most people realize, myself included.

what it is: server-rendered PHP with lightweight client-side directives. you add data-wp-* attributes to your markup and a small store in JS handles reactivity. full-page caching works natively. no hydration mismatches. it's WordPress-native interactivity without reaching for React.

it's not replacing React in the editor. React still powers the block editing UI. but for the frontend; product filters, cart updates, interactive elements; IAPI is now the standard path forward.