agentic software engineering. i run a multi-agent dev org: 30+ always-on bots that ship code. I move fast and share what I break
Joined September 2021
- Tweets152
- Following108
- Followers27
- Likes719
Pinned Tweet
When I was a student at VCU, I used Little Ram Pantries when I needed food.
Now I'm the lead research engineer on the tech behind them. 16 pantries have sensors and shelf cameras across two campuses, so students can check what's stocked before they walk over. And we have much more insight into the why and how behind pantry usage.
About 13,200 visits since 2021 and only going up.
VCU’s Little Ram Pantries have received upgraded technology to help address student food insecurity.
New sensors can show the inventory of each location and help monitor usage while preserving privacy. news.vcu.edu/article/little-…
Joshua Hayes retweeted
Look guys if you’re building something on a cloud platform and you don’t want to get owned through a platform compromise at some point you have like 4 real options
AWS
Azure
Cloudflare
GCP
You should not trust key infra to a neocloud
Neoclouds may get you faster performance on something like S3 but they just can’t match the availability, durability, and security of a “real” cloud
They don’t and can’t design for as many nines
an agent rewrite failed the same check three times. my agents automatically stopped retrying and made it a lint. now the mistake can't ship, no matter which agent writes it. that’s the power of pstack’s principals
154 prs merged into jstack in one week, mostly by agent crews.
the next bottleneck wasn't speed. it was two crews editing the same file.
so one bot now owns the map: every file in the repo belongs to exactly one feature. 998 files, 0 unowned, 0 owned twice. a brief that crosses features can't go out without a label saying so.
parallel agents need borders before they need more agents.
Joshua Hayes retweeted
When comparing OpenAI and Anthropic models, you shouldn't be using GPT-6 Astra against Opus 5.5 just because Opus is that good. You're comparing models from different tiers.
Astra's competitor is Fable, and Sol's competitor is Opus. Simple as that. We just need to admit that OpenAI has a gap to close, because Anthropic's tier below Astra is already giving it a serious challenge. And I believe that's about to change.
anyone else fighting approval prompts between their own Grok Bots?
my setup: an expediter that owns the queue, experts that research the queued work, and an orchestrator that hands work off to T3.
when i tell one bot to do something and it passes that to another bot i own, the second bot asks me to approve again in its own chat. one instruction from me turns into two or three approvals. it hits hardest when i'm spinning up new bots or moving lots of work across the team.
what i want: mark specific bots as trusted relays. if my expediter passes along something i asked for, the receiving bot treats it as my approval. opt-in, per bot.
the extra prompts don't make me feel safer. they just slow down the multi-agent setup i'm trying to build.
told my agent org "go with the defaults, show me a recap in the morning. don't stall on me" at 11:21 pm.
by 1:22 am it had answered 11 design questions my crews were blocked on. most came with a written undo, usually one commit.
the stuck crews and a merge hold still paged me, because those had no safe default.
if you want agents running overnight, every question they ask needs a default and an undo. then "go with the defaults" is an instruction, not a gamble.
My agents are no longer the bottleneck to progress, my PC’s CPU is. Guess who is going to Micro Center today!!
Joshua Hayes retweeted
i remember early in my career wondering why most people missed out on building early internet products
but i look around now and i get it
so many people fiddling around with "my custom agent setup" and so few using this miracle of a tool to actually do anything
Tokenmaxed my Claude Max subscription twice in a week. Can't believe this would've costed $4,488.03 without a subscription, I guess I need to get another one...
For years, spending 5 hours to automate a 5-minute task was a running joke. With AI, it takes 5 minutes.
The engineers who always tinkered now ship the most, and the ones who aren't experimenting are getting left behind
After watching a flying car and realizing I could be more ambitious, at 1 AM last night I gave a single instruction to a single Grok Bot. “Take every default, build everything you can, and set aside only what needs me. No spending, no prod deploys, and nothing that can't be undone.” Then I went to sleep.
That one bot splits my brief by repo and hands each piece to a long-lived home agent. Each home writes its own briefs, runs crews of coding agents, puts one reviewer on every PR, and merges once CI passes on my own hardware. Specialist bots handle product vision, market research and design. A message bus tracks every brief, approval and stall. Only the front-door bot is allowed to page me.
By 8 AM, 9 PRs had merged green on dev for Winnifred, a shared care record for pets on daily meds based on my own needs with my own dog. They added one-tap dose logging with Undo, WCAG 2.2 AA accessibility, a design system, hardened auth, a new data layer, ops alarms, pricing and a landing page.
Market research landed mid-build, and the front door turned it into briefs that moved reminders and insulin up the list. When the disk filled and the bus timed out, the agents recovered with little help from me.
I talk to one bot. It runs the org. The skills and playbook came from @poteto's pstack, and the system makes itself smart and smarter every day
Joshua Hayes retweeted
Accenture has told employees it will only fund 100 rows of excel per month. Meaning that employees who often use as many as 1,000 rows are scrambling to find ways to keep access.
New: Accenture has told employees it will only fund $29 per month of Claude use. Meaning that employees who often spend as much as $1,000 a month are now scrambling to find ways to keep access to the tool. A sign of the continuing crack down of AI 404media.co/accenture-tells-…
Killed a project and 4 Grok Bots today.
Deleting them was easy, but the other bots still had them in memory, plus old ids in their skills and stale files on disk. Clouding context and resulting in misinterpretations.
So I asked @poteto's dr eggbot to build me a Janitor. Every week it sweeps memory, skills and files for anything dead. Feel free to use it yourself! x.ai/bot/dEgnf8vpLirfZiR3xEa…
