Posts tagged: agents Archive

💭 A love letter to Pi | Lucas Meijer

2026-04-15T16:12:33Z

I hate how he called out terminal user interfaces as shit… then proved web interfaces to be superior. Damn him. I love working from my terminal, but having ai prove itself through html reports including video, image, metrics, charts, and text is goated. Rethinking yourself has the bottleneck not the orchestrator feels real. Validating the work is hard, theres a shift right now and everyone is trying to figure it out. Lucas’s technique is a little bit of be lazy and tell it to prove itself to you, so as you juggle your 15 agents you have a nice report to read.

Note

This post is a thought. It’s a short note that I make about someone else’s content online #thoughts

💭 How Claude Code’s Creator Starts EVERY Project - YouTube

2026-04-15T16:07:26Z

This is a really good guide, with quite a few good nuggets. I need to try deleting my AGENTS.md and rebuilding it from scratch more often. I liked how he talked about having agents prove their work and tell them up front how they will be judged. What I didn’t care for so much was the feeling that a lot of the rules go in markdown, thats not a rule, thats a suggestion. Rules should be deterministic. They should be tests and linters that ensure they are followed. Suggestions are good, but dont trust the agents to always follow them. And don’t trust that they wont change your rules, keep them honest.

Note

This post is a thought. It’s a short note that I make about someone else’s content online #thoughts

Agents Are Here

2026-04-10T16:11:37Z

🌱 This post is still growing

Late last year I started writing I'm Out On Agents. Agents sucked, the models were good, but there was still something missing between the harnesses and the models. They could write good code, they could do some debugging and exploring, but they were too good at fucking up the whole project to be useful. They could crank out Green Field POC’s like nobody’s business, but they created so much mess in brown field projects that it was easier to chat and edit yourself.

f91a8893-b1ba-422a-9390-18de5034483c.mp4

The Beautiful Glitch - Gemini

The Inflection Point #

It’s very well agreed on that the inflection point for most people happened with Anthropic Opus 4.5 in late Nov 2025. Early adopters probably noticed right away and shouted from the rooftops how good it was. But we’ve all heard that developers have 6 months before ai writes all the code for years, so this felt like the rest of the noise.

Hitting the December slowdown many of us hit code freezes at work. We completely disconnect from work for the last Week and come back in January. During this time, its very common for us to try out new tools, new techniques work on side projects, create POC for that thing we never have time for. While it looks like less features coming into the apps we support this is an important time for us to explore and reflect.

Anthropic was super generous with a free tier giving out huge levels of free tokens at the time. So many of us laughed and threw it at our side projects expecting the normal garbage output, but maybe some good ideas to come from it. But that’s not what was happening anymore, somehow these agents do some real work, follow plans and stick to scope really well. And if you lay out a big enough plan they tended to keep cooking and completing features.

January 2026 #

flu season

For me and many others around the country we were hit with a rough flu season, kicking in and draining us mentally for a good month or so. I wanted to work, I was getting excited about some projects and wanted to get them going, but I was constantly wiped and had no capacity. I had no ability to think on complex tasks, I was coughing all the time, just trying to survive, but yet wanted to do something. I started doing some small clean up, some work on side projects.

At this point I was still afraid of really letting agents cut loose on something meaningful, something that users depend on.

February 2026 #

what just happened?

yes or –dangerously-accept #

March 2026 #

April 2026 #

Closing the editor #

b289fda3-35e7-4b2f-aa21-6601c57d937a.mp4

Closing the Editor - Gemini

Thinking about ai productivity again

2026-03-16T21:01:44Z

Thinking about AI productivity again. It’s allowing massive amounts of work to get done, to levels that humans cannot physically type out in some cases. But not all of this work is necessarily high value work. Right now I’m working on one of the biggest PRs to an internal cli library. Probably the largest PR I’ve ever done professionally. It touches all of the cli, refactors every command, reaches into the business logic layers to drive deeper separation. I reaches into the common layers to drive consistency. It ensures that every command (50 or so) has similar flags, supports –plain, –no-color. It specs out contracts to ensure that data goes out stdout, any extra goes out stderr. This makes everything unix pipe friendly. There was quite a bit of research and prep that went in, that turns out to already be distilled down into clig.dev. The point is that this is all good work. It will make the product consistent, repeatable, expected, and most of all boring. Most of the time, it will just work. Since we did it ahead of a lot of other agentic work on the product its establishing good patterns for the product moving forward. But its low value work. We wouldn’t have likely put humans on this work wholesale and fixed critical paths as they came up. Its not cutting cost, selling more product, or driving critical business decisions. Yes it’s worth it now, but it would not have bee in the past.

💭 OpenClaw: The Viral AI Agent that Broke the Internet - Peter S...

2026-02-16T10:19:08Z

Such a good interview @lexfridman is such a talented interview. It’s so cool to see the other side of this. For weeks we’ve heard about the story of the name change, we’ve seen everyone shitting on the security model, buying up all the mac minis in existance, fear mongering not to install this thing. @steipete.me has such a cool story from the beginning talking about making this thing fun and exciting. Giving it a personality that is not “You are absolutely right”. The story of changing the name twice, and getting pwnd on every step the first time and nailing it the second time is incredible. Dude is having fun trying to make the thing he wants in the world exist.

Note

This post is a thought. It’s a short note that I make about someone else’s content online #thoughts

Pm Not Babysitter

2026-02-06T14:04:25Z

Stop babysitting your agents, treat them like a real team and they will reward you.

Back in December I saw theo make a comment that code is now cheap, its the run rate of models, He quoted a study, not sure that he fully even believed it, but it claimed that the average developer after all meetings, training, emails, planning and extra shit in their day averages out 10 well tested lines of code per day. Opus 3.5 made him 10k loc (lines of code) that day.

We have all agreed for decades that lines of code is not a proxy to productivity or quality. Often more code means more risk, more review, more infrastructure. This has become MUCH different. Lines of code are still far from any sort of good metric. That aside, your agents are not doing 10k lines with you babysitting them, and in fact its very likely that the product quality is MUCH worse as you babysit them.

You need a tool for planning and tracking, otherwise you are playing babysitter rather than Product Manager (PM).

Agent Management Is Exhausting

2026-01-26T14:38:35Z

The state of development in early 2026 is all wrapped around learning how to manage many agents running in parallel. Everyone’s trying to figure out the workflow.

The Plan Is Everything #

The secret I’ve discovered is a good, well-defined plan. This could be a markdown file or a GitHub issue. Agents are actually great at writing these for you. They’ll include reproduction steps, outline changes needed, and structure the work.

This is your opportunity to step in. Read the plan. Look for hallucinations. Spot where it’s going off track. Edit the plan before the agent starts coding.

I had one today where it laid out reproduction steps beautifully, but I could add context about network requests that completely changed the approach. This editing phase is what most people are missing right now. Skip it and you’ll watch your agent solve the wrong problem with impressive efficiency.

The Pace Problem #

Here’s what nobody warned me about: managing these things is exhausting.

Depending on the day, agents move so damn fast. I can barely research, find, and raise issues as fast as Claude can implement features and fixes. It’s like trying to speedrun a Minecraft seed when you just figured out how to craft a pickaxe.

A Different Kind of Work #

This stretches a different part of my brain than I’m used to using. I’m learning new skills around:

Issue tracking and management at high velocity
Knowing which models handle which tasks best
Spotting when a session is about to go sideways from context bloat or bad compaction

I had a session yesterday where the context got poisoned with a wrong assumption. The agent spent 20 minutes building on that false premise before I caught it. That’s 20 minutes of perfectly executed code solving the wrong problem entirely.

Worth It, But Hard #

I’m not going to hype-bro you and say this is easy or that you’re wrong to be skeptical. If you have the opportunity to work with agents, it’s worth learning proper techniques.

It’s not easy. It requires actual management skills, not just prompting skills. And yeah, I’m tired.

Context Is King

2026-01-13T08:28:23Z

A new approach to agentic workflows.

This is probably news to no one else, I’m sure I’m behind on this one. You can’t one sentence prompt and expect to get what you want.

Agents.md #

Writing Complete Prompts #

I'm In On Agents

2026-01-11T12:02:15Z

It’s the start of 2026 and agents are getting a lot better than they were. I’m using opencode at home, free mode with Zen and big pickle. At work I have access to a wider variety of models including what seems to be the gold standard 3 from anthropic opus, sonnet, haiku.

Note

I’ve sat on this post for over a month, refactored it a few times, things feel like they are moving so fast that its hard to keep my feelings straight.

Things are changing #

Around Aug 2025 I wrote I'm Out On Agents. I saw others in the space having such great success I gave it a solid shot, but found it to egregious edit more than I asked, make massive unneeded changes, and make more small bugs hidden in the details than was worth it.

This was just after the tipping point where not just hype bros were making the switch, but people I work with, people I trust and have made really good Software in their career. People like the creator of Flask Armin Ronacher. People who are really good at writing code and do not need it to do the work were finding themselves more efficient working with agents than making edits by hand.

Planning #

Currently as I see it there is no free lunch. Software engineering is still very much needed, but the work is switching. The models have been good for awhile, they don’t seem like they are getting leaps better with each release like they were for so long, no matter how much t3.gg hypes up the new models. The tools feel like what is really making the difference right now. The ability to take short human prompts figure out what context is needed, getting that context, and keeping it in the context window is getting better and better making agents useful for more things.

These improvements in tooling is making our job look different. Knowing syntax and how to write clean efficient DRY code is less and less of the job every day. Learning how to create detailed plans with the help of agents, editing them, knowing when they are good enough that you can be confident in an agents ability to execute on them. Knowing how to balance the context is becoming the job. Moving forward its likely that models can handle more context and this will look different, and change, but planning and context management is a key part of the job moving forward.

Context is king, good plans are paramount, syntax barely matter.

Issues #

I'm Out On Agents

2025-08-27T10:10:51Z

Its the year 2025 and we are only a few years into having 6 months to live before ai takes our jobs, and the big push right now is agents, managing agents. I will fully concede to I’m not doing it right, or a future state gets better than where we are right now, but right now they kinda suck.

Transparency

I’m sitting offline right now as I write this, These are my feels, no research, no links, no ai, just vibes.

Chat #

Chat is what really kicked off ai uses and goes back as old as computers, but it always sucked. Then chatgpt rocked the world with the biggest launch day in history and showed us that it could actually be pretty good. Unethically trained on everything they could get their hands on, burning cities worth of electricity to train, and keep training to stay ahead of the competition. It does a damn good job. There are tells, and if you see enough of it there is a lot that turns to slop, but if you had never seen it before, there is no way you would assume that it was not a computer.

It does a damn good job at being average, it can do what seems like everything not related to security and authentication at a pretty average level. This is its super power. Whenever you are trying to bridge between something you know and something new, you can get a pretty good answer, and likely spot the bs in what you know.

Agents #

Now that our models have gotten better, hardware has gotten bigger, better, and a lot more of it, we can really expand context windows really wide. With that brought the use of agents, these tools get context from sources on their own and often are given read/write access to your computer. Depending on the model these things will branch out to make small changes that look no worse than a formatter on every goddam line of your codebase. Except they are not a formatter, they are not backed by ast checks. They do not have any guarantees.

They take the fun out of creation.

The emphasis now becomes on the code review. All you have to do is ask it to makes changes for you. Bring in your expertise of what changes should be next, or even just punt to asking it what comes next. I’ve yet to talk to someone that is diligent enough to read everything it spits out in excruciating detail to the point that it does not cause significant issues.

You see here is the thing, its average. With a little bit of context it can do average work. It is not an expert. On the surface this feels fine, making crud endpoints has been done for decades, and average developers crush these things every day no problem.

It’s Average at best

You know what is not average? Your knowledge of the use case you are solving. You may not feel like an expert, but given that there are likely about 5 people working in your codebase, you are a fukin expert at that thing. The average person off the street takes time to onboard, often months, or years for someone to really understand the business you are working in. These things don’t have that.

I’m not letting it in shit that I care avout #

I was early to the game to codeium, even used the predecessor for awhile, I was early to chat gpt, I was early to windsurf and the possibilities that agentic ides brought. I will be late on letting agents touch my production code bases. The few times I’ve tried for changes that seem easy enough, but more work than what I want to do at the time, I’ve regretted it. It’s only left behind a mess that it cant deal with anymore, runs in circles trying to solve any problem, and I’m left with shit that feels like a house of cards that breaks anytime you touch it.

I will continue to POC #

As much of a mess as these things, make they are super useful to vibe code ideas, move quickly and try different approaches to a problem. They let you make a proof of concept that you can get in front of team mates, bosses, or users.

Honestly I still prefer the chat interface. It feels like a nice balance of using my brain, and knowing that I am punting on something. Do I need to know the whole ffmpeg interface to grab a thumbnail for my webapp, no, would I actually like to learn it someday, yes, but I don’t need to right now I just need a goddamn thumbnail.

For the most part I am still slotting these things into the codebase myself, but occasionally I have it do an entire module, and even more rare I pull out agents and have it do all the work. The further away from the code I get, the less I care about it and just want it done.

I will laugh at this post in 2 years #

The way I see it agentic coding is not here to stay, it sucks. There is one two ways to go this was a blip in the radar that we laugh at for giving ai companies all our data for free so they could fuck up our products. Or they actually get pretty good and we all become architects that peek at the code if we really need to. I think for the second to become true its going to take a lot of time, consumer hardware will need to catch up, local llms will need to work a lot better than they do now. Even if we still need to call out to the big boys for some heavy work getting the context right for them locally would make a huge difference. Currently local llms are too slow and dumb.

So will this bubble pop or explode, we can only wait to find out.