It is good to look at and learn from agentic orchestration, but most important to understand the purpose behind the approach. Strictly quantified evals leads to more deterministic understanding of individual prompts. It is a way to build strong building blocks which is important in a large system with lots of other contributors and moving parts. It allows incremental improvement and % success rates and all that.

On my pet projects though, I avoid agentic orchestration. I'm cheap, and most orchestration is deterministic - I start here and want to get to there. Why pay for AI to manage a known process with gates and triggers? That makes it much harder to reason through possible outcomes. It may be more flexible and powerful, but that is the "multiple camera angles" concept, where you build defensively "just in case". If I know what I want the next step to be, just do that directly. I prefer to start with a clear DAG and build on top of it, keeping AI calls narrow and compose outputs not processes.

Yes, you can go to town with agentic processes but eventually you hit the wall of understanding. If you lose the thread of what an agentic process does and how it makes decisions, you lose the ability to find shortcuts. At the root of all hacking is curiosity to know how things work. Taking things apart and manipulating the underlying system is exactly how all of these new technical advancements happen, not by following "best practices" that didn't exist 2 years ago and were built to maintain a manageable engineering department more than maximize AI value at minimal cost.

I think right now we are waiting for the Morris worm (https://en.wikipedia.org/wiki/Morris_worm) equivalent shock to the system, but it is likely to be much, much worse and much more specific. I expect something that will make DOGE stealing SSNs look kind of tame. Something like every private GitHub exposed, every Visa card data and history exposed, every Mac injected with a rootkit, etc. It's like waiting for the plot from Sneakers to manifest.

For all the security we have built over the last 50 years, it has been impossible (or nearly so) to lock down any web-accessible content. It is a structural issue at a certain level of complexity, the surface area is just far too wide for any focused effort. Aside from direct 0 day vulnerabilities in software there are vulnerabilities in core libraries, frameworks, CI/CD, cloud services, hardware bugs, gaps between services, permission vectors, etc.

The U.S. has relied on the legal system to allow our insane credit card system to persist, where security by obscurity (knowing someone's CC#) is the main deterrent to abuse. I need a complex password to access any website, but CC#s are flying free. I think the combination of easy worldwide vulnerability scanning and U.S.'s focus on pissing every country off is going to lead to significant and unending asymmetrical warfare. If our gov't has been co-opted by big business, big business is going to become the target. As we have seen with Iran with Hormuz and Ukraine with drone strikes, it isn't so hard for small countries to fuck up global systems.

We are entering a 90s-style phase where any script kiddie can cause massive disruptions. Trump likes to threaten NUCLEAR but security issues could potentially cause even more death and destruction - overwhelm the energy grid, open dams, crash air traffic control communications, etc. There is lots of concern over the oligarchy owning AI and keeping it for themselves, but the more immediate risk is that any country can potentially lash out with disruptive actions.

There has been a retreat from globalization since COVID. I wouldn't be surprised if that extends to global internet communications as well. Internet traffic between countries might soon be severely restricted, that's the last line of defense we actually have if this goes as badly as Anthropic is implying.

Most telling: I mistakenly filtered against all jobs in SF on LinkedIn and I was a few pages in before I noticed. There were just a few lawyer jobs in the mix but literally every other job was AI/backend/product engineering roles. Meanwhile, I know an experienced EM from enterprise SaaS who has all but stopped looking for work.

In the 1800s and early 1900s, complex sentence structure signaled intelligence for both the author and reader. It was a form of entertainment, in a way, when books were few and nights were long. Try reading Henry James for an idea about what this looked like in practice. Shakespeare is another obvious example of "heightened language" besides archaic words the play are written in iambic pentameter and the spoken text is far from natural (yet incredibly precise).

For example, a customer service playbook may have certain ways to handle different user problems, but that breaks down as soon as there are complications or compound issues. But an LLM with the ability to address individual concerns may be able to synthesize solution given fundamental constraints. It's kind of lile building mathematics from axioms.

This could likely be extracted much easier now from the new code, but imagine API docs or a mapping of the logical ruleset with interwoven commentary - other devtools could be built easily, bug analysis could be done on the structure of rules independent of code, optimizations could be determined on an architectural level, etc.

LLMs need humans to know what to build. If generating code becomes easy, codifying a flexible context or understanding becomes the goal that amplifies what can be generated without effort.

For $20 a month, I can plan and implement thousands of features using Composer 2 or Auto with Cursor. The usage limits are insanely higher. Yes, the depth of understanding is not Opus 4.6, but most work doesn't need that. And the work that does need it I pass to Claude.

I can code 8 hours a day using LLMs as my primary driver spending just $40 a month.

Remote: Yes, or in-office/hybrid in the city

Willing to relocate: No

Technologies: React, Next.js, JavaScript, Python, Django, Flask, Postgres, MySQL, Redis, PHP, Docker, Github Actions, Claude Code, Cursor, REST APIs, OAuth, LLM orchestration (OpenAI API calls, LangGraph, LiteLLM, SDD, GSD, etc.) Plenty of off-resume exploration of other technologies (Node/Express, Ruby, Mongo, CI/CD tools) and comfortable picking up languages and tools on the fly.

Portfolio: https://yagmin.com/

LinkedIn: https://linkedin.com/in/jyagmin

Ex-founder/CTO (built a team, grew to a top 3000 website, successful exit) and 20 years experience as a fullstack product engineer. 0-1 building for both B2C and B2B, worked on data-heavy enterprise SaaS but most of my experience has been seed to series B startups. Worked as an IC the past 5-6 years in founding/senior eng roles.

On the AI front: In the last year I've built a production user-facing AI app and also an open source context management harness for Claude. Up to date on the applied AI landscape. I've done a lot of noodling with different tools and approaches.

Looking for a role where my product and tech experience can make an impact. I'm deeply invested in applied AI and excited about how AI is reshaping how products are built and served.

Cursor also has an interesting Debug mode that actively adds specific debug logging logic to your code, runs through several hypotheses in a loop to narrow down the cause, then cleans up the logging. It can be super useful.

Finally, when making peecise changes I can select a function, hit cmd-L and add certain ljnes of code to the context. Hard to do that in Claude. Cursor tends to be much faster for quicker, more precise work in general, and rarely goes "searching through the codebase" for things.

Most importantly, I'm cheap. a If I leave Cursor on Auto I can use it full time, 8 hours a day, and never go past the $20 monthly charge. Yes, it is probably just using free models but they are quite decent now, quick and great for inline work.

When AI can write convincingly enough, it is basically a honeypot for human readers. It looks well-written enough. The concept is interesting and we think it is going somewhere. The point is that AI cannot write anything good by itself, because writing is a form of communication. AI can't communicate, only generate output based on a prompt. At best, it produces an exploded version of a prompt, which is the only seed of interest that carries the whole thing.

Somebody had that nugget of an idea which is relevant for today's readers. They told the AI to write it up, with some tone or setting details, then probably edited it a bunch. If we enjoy any part of it, we are enjoying the bits of humanity peeking through the process, not the default text the AI wrote.

What stuck with me is that they found that by elongating the yellow stick, the baby birds would yell even harder than when their mother was there. In other words, are instincts amd impulses are imprecise and can be manipulated.

This is not a new thing, though. In some ways, this is something art has long manipulated - no love is more tragic than Romeo and Juliet's, for example.

If you want to see a big file, I have a direct link to load the Blocktronics WTF4 ANSI which is over 4000 lines: https://sure.is/ansi/?wtf4=1