As an extra bonus, beads UI helps me interact with the beads database like with a Kanban board. The UI is pretty basic, but enough to click around and browse the tasks, update their descriptions when needed, etc.

🔗 https://github.com/mantoni/beads-ui

A very interesting talk on how to use AI agents on large codebases in a way that does not generate slop and produces more predictable results. Matches a lot of what I've been doing over the past 6 months.

🔗https://www.youtube.com/watch?v=rmvDxxNubIg

If you’re interested in building custom agents for your business (or your personal needs), here is a really helpful overview of typical problems you may need to solve: from skills support, to sub-agents, to code-driven workflows.

🔗 https://lethain.com/agents-series/

https://addyosmani.com/blog/next-two-years/ - really interested blog post on the next couple of years for the software engineering industry. I liked how the author tried to extrapolate the trends from the past couple of years and then looked at it from the point of view of different kinds of engineers and how they could fit into the new picture.

https://antirez.com/news/158 - very good article from the Redis author on his view of the current situation with AI models and how they are reshaping the coding industry.

https://lucumr.pocoo.org/2026/1/18/agent-psychosis/ - an amazing post from Armin on the mental problems people experience when starting to build with AI agents. I can 100% relate to a lot of the issues he is describing, including the dopamine loop issue I'm still struggling with (the "one more prompt" situation at 2am after a 12-hour non-stop building session).

https://cannoneyed.com/projects/isometric-nyc - A great example of a hobby project made possible by modern AI coding tools and image models. A great article on how the author approached the creation of a pretty complicated piece of software using modern AI coding agents.

> As a software developer, I think of a million little tools I’d like to have but would take a day or a week to build. With Claude or Cursor, I can whip them up in 5 minutes. This is absolutely transformational - it’s like having an infinite toolbox.

This resonated the most.

https://openai.com/index/unrolling-the-codex-agent-loop/ - extremely detailed write up on how Codex coding agent from OpenAI works internally, specifically how the agentic loop is structured, what information is sent to the LLM and when, etc. Fascinating for anybody building their own agent!

https://engineering.block.xyz/blog/ai-assisted-development-at-block - tons of details on how they use AI agents at scale at Block (had no idea Square changed their company name 😅). A lot of actionable insights for anyone interested in bringing AI into their teams.

https://blog.julik.nl/2026/01/go-fast-and-go-alone - an *amazing* blog post on how we should be changing our software delivery process in the world of LLMs. HIGHLY recommended reading for anybody trying to apply these new tools in a team setting. 👌🏻

https://shumer.dev/something-big-is-happening - An amazing article on the progress of AI tools and how it has started to drastically to change our industry. An article that you can share with your friends and family who may not be up to date with the recent developments in AI and are still skeptical about what is coming.

Loved this quote: "This might be the most important year of your career. Work accordingly."

https://www.anthropic.com/glasswing - Amazing initiative from Anthropic. Patching the world before releasing the model that poses security risk to the global economy. Brilliant marketing as well!

https://x.com/addyosmani/article/2052124873208799378 - really good article on the problem of Cognitive Surrender to AI. Matches my experiences perfectly. Just like with technical debt, it is helpful to keep it in mind while working with AI agents and try to be deliberate about the things you surrender on.

https://desloppify.it/ - really interesting structured approach to reducing the levels of "AI smell" (aka slop) in your codebase. Have been using it on our projects for about a month and even if not all of its recommendations are spot on, a lot of them are very much spot on. And the sloppiness metrics are helpful as well.

https://nolanlawson.com/2026/05/25/using-ai-to-write-better-code-more-slowly/ - very good little post that matches closely how our team uses Codex a lot (applying a similar process to each PR).

I shipped a small open-source tool today: Agent Secret.

The problem it solves is very specific, but I think more people who use coding agents heavily will run into it soon.

As agents get more useful, they need access to more real credentials: API tokens, deploy keys, cloud accounts, database URLs, etc. We moved this stuff into 1Password, which is a big improvement over random .env files. The op CLI plus Touch ID is already pretty good.

But when you have several agents running in parallel, a random Touch ID prompt asking for 1Password access stops being meaningful.

Which agent asked for it? Which secrets does it want? Why? What command gets the values after I approve?

Agent Secret sits between the agent and 1Password. Instead of calling op directly, the agent asks Agent Secret to run a specific command with a specific list of 1Password references and a reason. I get a native macOS approval prompt with that context, then approve or deny.

We’ve been using it daily for months now. On busy days it handles dozens of approvals across personal, employee, and shared team vaults. At this point, raw op prompts feel like a step backward.

It is very much v0:
- macOS only
- 1Password only
- single exec commands only
- no secret writes yet

That is intentional. I wanted the first version to solve the approval-context problem cleanly.

Site: https://agent-secret.sh
Repo: https://github.com/kovyrin/agent-secret

https://antirez.com/news/169 - a really good post from the author of Redis on how since 5.6 we're basically entering a new era of software development where reviewing code line by line is getting a lot less meaningful.

I tend to agree with him on this 100%. Since 5.5, while reading all the code, I've been mostly asking it to change things I did not like, but that would 100% have worked as-is. Just taste stuff. And 5.6 started doing it better (or following existing conventions better) and now I find myself at a loss trying to find anything there in its diffs.

With 5.6 I slowly started shifting towards the model he is describing as: asking how the code works, reviewing that, challenging assumptions, doing QA, etc instead of reading it all line by line like I've been doing for the past year.
Exciting times ahead!

https://blog.exe.dev/claude-is-not-a-compiler — I tried his differential spec analysis approach last night (https://commaok.xyz/ai/differential-spec/) and was very impressed with the result.

It launched three sol xhigh agents for me, built three complete prototypes, and then launched two more agents to analyze them - one focused on the architecture, while the other looked at user-facing behaviour, API contracts, etc.

In the end, the process found a ton of gaps in the plan, as well as many places where the agents had stumbled, which it also documented in the plan. Highly recommended!

https://blog.cloudflare.com/kitesurf/ - really fun read on how Cloudflare folks have used AI agents to build an agent-focused browser implementation on top of v8 and Cloudflare workers. Really shows the value of real software engineering in the age of modern AI agents!

https://www.youtube.com/watch?v=87DyyMV0kCY - an insane glimpse into the future of offensive cybersecurity... As someone who has been a part of defensive security efforts at multiple large and small companies over the past decade, I am 100% certain that none of us are truly ready for this. 😰