Tonight Iโll sit down with @pmdroid to talk about agentic engineering in a multiplayer setting and how heโs shipping and building!
join us live at 23:00 Rio time (18:00 PT):
https://t.co/oHUMfgH5j9
You don't want to miss Monday night's stream, I'll be sitting down with @pmdroid to chat about agentic engineering and the cool projects he's building around it!
(may or may not be interrupted by baby feeding and burping, you have to show up to find out!)
last year i did almost all my work in mobs or pairs. one screen, everyone knew what we were building and why
this year everyone got their own agent and headphones. faster, but alone
i want the team running agents together again. so i'm building Paws ๐ง
https://t.co/86LMBxxIUL
last year i did all my work in mobs or pairs. this year everyone went solo with their own agents and i don't love it
this is the direction i want: one shared context, one board, agents that coordinate with each other. working as a team again ๐ฅ
https://t.co/jNE36huBtX
No. We've got to stop applying old human-centric patterns to working in the new era.
delete all your wikis.
have an agent answer questions from source instead.
at spawn we have:
- zero dashboard tools
- zero docs
- we do not make or use summaries
everything we do is primary sources + intelligence
Nothing to go stale
Nothing to organize
Nothing to navigate
Always exactly up to date
Perfectly personalized
The answers get better with every model
If we want to see a chart, we ask for it, and we have it in a minute
If we want an answer to a question, we ask it, and we have a to-the-minute correct answer
The homie Pascal Matthiesen (@pmdroid) says your agent doesn't need more docs.
Spicy take but he thinks you should kill most docs (unless they reference code you don't control) and docstrings. My exception is agent tools, where the docstring is the only documentation the model never skips.
He also shares a prompt to check whether your docs match what your code actually does.
https://t.co/KPBV8Hd4KE
Why play the maze when you can over-engineer the maze? ๐ค
Gave it a model that picks a direction and says how sure it is. I'm now watching my website play itself.
https://t.co/D822kX9YDc
I turned my personal website into a maze. Finally, getting lost on a website is a feature ๐
A silly idea this morning + a few hours of Kyoto Agent doing its thing.
Come get lost ๐
https://t.co/Yay8VgShFP
@GitHubNext@MSFTResearch Love this. I've been building the lower-tech version of the same idea ๐ซ
Closeout is an open spec. A file in the repo says what has to be verified before agent work counts, and the agent has to bring the evidence.
Not proofs, just receipts. https://t.co/CAMgslBWOJ
Agent: "Fixed the checkout form โ "
Me: "Show me."
Agent: ๐พ
Before/after, screenshots, video, all in the PR. That's what a receipt looks like.
https://t.co/fA3T1LVQ9b
Amp says mandatory code review is dead. Fine. ๐ซ
But "nobody reviewed it" and "nobody can prove it works" are two different sentences.
Skip the reviewer if you want. Keep the receipts. ๐งพ
How are you proving agent work is done?
https://t.co/CAMgslBWOJ
I built a fake Linear so my agents can mess up issues without anyone at work seeing. ๐ซ
Finear runs locally on SQLite. It works with the real Linear SDK, OAuth, signed webhooks, and agent sessions.
https://t.co/BopHwVRQOy
Your AI agent says it fixed the UI.
Do you believe it? ๐
I built Pawprint so it can bring receipts: screenshots, videos, and before/after comparisons, linked right in your PR.
โDoneโ looks better with evidence ๐พ
https://t.co/fA3T1LVQ9b
Example: https://t.co/0ywXBan54f
Remember when the best way to get better was pairing? Two people, one problem, everybody learns ๐ค
Then AI agents showed up and we all went back to coding alone. Six terminals, me approving every tool call, nobody else in the room ๐ซ
I'm building it the other way: the epic lives in Linear, agents do the work, and when one gets stuck it asks the team in the issue. Whoever knows answers, and everyone sees the answer.
Agents should bring back pairing, not kill it. Are yours a team sport yet? ๐ค
@Joshua_WD Kyoto fixes the independent review findings and retries on its own while the retry budget remains. Once that budget is exhausted, it blocks further attempts, completion, and PR creation, then hands control back to you.
I built another coding agent. Yes, another one. ๐
Meet Kyoto Agent: less watching an agent work, more focusing on what youโre building.
Explicit permissions. Clear results. Checks behind โdone.โ
Named after my golden retriever.
https://t.co/3Vcsa071hN