Thomas D

@tomasruizlopez I’m just sayin: it’s much easier to review the code if you have an understanding of the general shape it should have. You should spend less energy reviewing code and more achieving alignment with your agents and coworkers.

@tthomasdd You may also be reviewing something you didn't produce. And we know that even if you achieve alignment there is no guarantee the produced code actually matches what you agreed with the robot!

nobody actually has things figured out btw

@pvncher ahh yes, the main purpose of sub agents

Even better than this - ask codex to use scout agents to set do the research first so you don’t clutter the main context window. Sub agents to resolve ambiguity, then you run

vibe coding allows "ideas guys" to bring their ideas to reality, revealing that their idea kinda sucked

Apparently, is that time of the year when people migrate from Anthropic to Cursor. Last time it took 2 weeks (July 1-July 15) to boomerang back. Any bets? In any case congrats, boomerang or not, the employee wins.

herdr hit 16k stars! which means more issue triage, support, and roadmap pressure, so i’ve been burning tokens left and right keeping up. applied to both openai and anthropic os maintainer programs a while ago. still haven’t heard back. once token-poor, always token-poor :(

@giuseppegurgone @pidotdev oh i understand, you mean no fixed personas.

@tthomasdd @pidotdev they are created on the fly by the main agent using the instructions in AGENTS.md, no agents files definitions.

I am experimenting with ephemeral subagents that PI (a custom extension) generates on the fly using a few lines of instructions from AGENTS.md The two agent definitions are in the screenshot.

i get asked what memory system i use often. my answer has always been: my codebase is my memory system.

@fchollet Hilariously, there is a post with similar engagement that states the exact opposite. Everything is slop.

The weak AI code gen we had until late last year was most useful to low-skill programmers -- it was raising the floor. It was essentially useless to high-skill programmers -- you could move faster and ship better code without. This has been completely flipped: the strong AI code gen we have now is *most* useful to high-skill programmers, while low-skill programmers are vastly underutilizing it or sometimes drowning in it. It went from a crutch to a power tool.

@dexhorthy dog fr i'm always like: ok say all that again but shorter and with simpler words and just give me the high level ideas and we can dive deeper into the details if necessary

codex prompting guide

@mikeyobrienv any way we can get rid of the emojis?

Agreed, though many of these go beyond the scope of a tidy subagent primitive. How many of these things does a fable/gpt-5.6 class model REALLY need? 1. Claude Code look and feel 2. Parallel background agents 3. Live widget UI 4. FleetView 5. Conversation viewer 6. Custom agent types 7. Mid-run steering 8. Session resume 9. Graceful turn limits 10. Case-insensitive agent names 11. Fuzzy model selection 12. Context inheritance 13. Persistent agent memory 14. Git worktree isolation 15. Skill preloading 16. Tool denylist 17. Styled completion notifications 18. Event bus 19. Cross-extension RPC 20. Scheduled agents 21. Model-scope enforcement

If you like pi-tidy-tools, you might like pi-tidy-subagents. Same minimal UX. Fan out tasks across model + thinking combinations. pi install npm:@mobrienv/pi-tidy-subagents github.com/mikeyobrien/pi…

@doodlestein Plot twist: this just iteratively creates more complexity than is warranted. Models can and will find every possible thing that is underspecified or not quite just right, until forever, when your 300 line code change ends up being 3000.

It doesn't seem to matter how smart an LLM is, this prompt basically always helps it find problems in its work: "Once again, check over everything again with fresh eyes looking for any blunders, mistakes, errors, oversights, omissions, problems, misconceptions, bugs, etc. Be SUPER thorough and meticulous!" Even after 3 rounds in a row! The harder the tasks, the more problems it finds. Doing this during the planning phase saves a lot of tokens and tears later. Measure twice, cut once!

@mikeyobrienv yeah but tintinweb's version is quite a bit reduced in scope from the original pi-subagents by nico.

@tthomasdd It's another full agent system for Pi. pi-tidy-subagents is deliberately not that. I intend to keep it as simple as possible. The focus is on UX and observability over more orchestration mechanisms.

it feels so good when the model asks: what do you want to do? a) ... (recommended) b) ... c) ... and your answer is not the recommended option. i'm like "ha! gotcha clanker!"

i just spend 30min debugging why cmd+click didn't open links (it hasn't for the past 2 months) and it turns out my footer extension was intercepting the clicks and never sending them to pi... terrible.

pi is the linux of agent harnesses

@mikeyobrienv How do you think it compares to github.com/tintinweb/pi-s…?

@tthomasdd pi-subagents is the batteries-included orchestration OS: roles, chains, async jobs, worktrees, review loops. pi-tidy-subagents goes back to basics and a focus on simplicity. Give models a powerful primitive and let it handle orchestration.