Manu Ebert

What startups are made of vs. what academics think startups are made of.

only in san francisco @hyperspell

@rlsayar Explain why you want to keep actions separated by role for agents? Why does it matter? Actions depend on knowledge, a more wholistic context means better actions.

@maebert I think you are confusing knowledge with actions In your example this shows (it's about asking a question) For knowledge it definitely makes sense but you want to keep actions seperated

Humans spent 30,000 years pretending to talk to god. And for some reason the ambition of most workplace AI agents stops at replacing Jake from accounting. If you're building agents, you're choosing between two futures. 𝗣𝗼𝗹𝘆-𝗮𝗴𝗲𝗻𝘁𝗶𝗰: One agent per function. A copy-paste of your org chart. 𝗠𝗼𝗻𝗼-𝗮𝗴𝗲𝗻𝘁𝗶𝗰: One agent that knows the whole company and adapts to the person/task. The poly-agentic model is a human-centric projection onto machines. We have separate departments because humans have limited bandwidth. A salesperson can't also be an engineer. Communication between humans is expensive and difficult, so we keep each person's scope narrow. But agents don't have limited bandwidth the way people do. With a proper memory solution, one agent can hold everything sales, engineering, and support knows simultaneously, with perfect recall. An omniscient, omnipresent agent that works on your company vision. Most workspace agent providers are still trying to push the poly-agentic model, one agent per function. 𝗕𝘂𝘁 𝘄𝗵𝘆 𝘄𝗼𝘂𝗹𝗱 𝘆𝗼𝘂 𝘁𝗮𝗹𝗸 𝘁𝗼 𝗝𝗮𝗸𝗲 𝗳𝗿𝗼𝗺 𝗺𝗮𝗿𝗸𝗲𝘁𝗶𝗻𝗴 𝘄𝗵𝗲𝗻 𝘆𝗼𝘂 𝗰𝗼𝘂𝗹𝗱 𝘁𝗮𝗹𝗸 𝘁𝗼 𝗴𝗼𝗱? At Hyperspell, our agent Cubert lives in our group chat. He doesn't have a "sales mode" and an "engineering mode." He knows what's going on across the company, and when I ask him about a customer issue, he pulls from the same knowledge graph as when @contextconor asks about a product roadmap question. 𝗕𝗲𝗰𝗮𝘂𝘀𝗲 𝗶𝘁'𝘀 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝗸𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲. 𝗜𝘁'𝘀 𝗼𝗻𝗲 𝗯𝗿𝗮𝗶𝗻. Build one agent that knows everything. Make it act differently depending on who's talking to it and what they need. It knows the whole picture, so it adapts. You don't need more agents. You need one that understands your company the way a chief of staff or trusted advisor does.

every non-founder at hyperspell makes the exact same salary. we don’t negotiate founders make less. i’m the lowest-paid full-time employee at the company we pay enough that nobody worries about money, but not so much that people optimize for it we’ve had ex-citadel, ex-google, a former surgeon, and engineers with 15+ years of experience take up to 60% pay cuts to join equal pay removes a surprising amount of politics. nobody spends energy wondering if they’re underpaid relative to someone else, negotiating for marginal bumps, or optimizing for promotion cycles the hiring question becomes: “do we want to work with this person?” the most valuable people are usually optimizing for agency, ownership, and who they get to work with

everyone at @hyperspell gets $1k/month they have to spend on making their life better use it or lose it, you can’t take it as cash people have used it for a home organizer, flights home to surprise their mom for mother’s day, even a cessna flight over canyonlands everyone posts what they used it for in slack and competes for creativity we ask a lot of each other. some team members took big pay cuts to join hyperspell because they believe in what we’re building the money matters less than someone telling you to take care of yourself

Artificial scarcity ⁦⁦by @AnthropicAI⁩ around Mythos is brilliant. Gets labelled supply chain risk. Hypes up “most advanced hacking tool on planet”. Puts US gov’t in a loose / loose position: deny the NSA access, or be called out on their bluff. axios.com/2026/04/19/nsa…

@t_blom @microsandbox Hey don’t tell me what a consenting human can do with his agents in the privacy of his own home

Are you rawdogging AI agents like a savage? You should use @microsandbox

@contextconor Good thing you didn’t get on a Forbes 30 under 39 list, or the con in Conor would have hit different

I was born to work on context for ai context conor, reporting for duty 🫡

Good job Hungary. Can we repeat the process with other autocratic populists elsewhere? I have prepared a short list.

@conor_ai Always wanted to host an award show for coding music called the Programmys

working this weekend listening to electronic music call it claude codechella

@nikhilaprabhu @conor_ai @realbenhorowitz The other one. The better one.

@conor_ai @realbenhorowitz ben horowitz lol

our founding eng @realbenhorowitz is so cracked he’s found critical bugs in 3 of our infra providers he’s the human mythos ai honestly our vendors should be paying us

Something @jack implies but deserves being called out: humans still think we need an authoritative, self-coherent source of truth on org-charts, projects, etc. We design processes and organizations under the delusion that we actually have that. Using AI to change how we work means focusing on what's relevant right now to you, not what's true for everyone.

@noor36758 rls

I'm a frontend developer, scare me with one word

I spoke to @CliveThomp15983 at @NYTmag about the future of coding, the depths of my prompt files, and the daily fight to get Claude Code to not embarrass me. nytimes.com/2026/03/12/mag…

@garrytan I remember fundraising a year ago and every investor was like “won’t people just use an MCP for this?” There’s not a single problem that MCPs solved that couldn’t be solved without MCPs already.

MCP sucks honestly It eats too much context window and you have to toggle it on and off and the auth sucks I got sick of Claude in Chrome via MCP and vibe coded a CLI wrapper for Playwright tonight in 30 minutes only for my team to tell me Vercel already did it lmao But it worked 100x better and was like 100LOC as a CLI

@the__shredator @conor_ai @hyperspell I love how a practice that is completely standard for professional TV people and literally called a prompter is considered weird for people who also have lots of other things to do than to pitch ideas on twitter

@conor_ai @hyperspell @maebert These weirdos can't even pitch an idea without reading from a prompt. Their brains have already been hijacked by an LLM

introducing agent-to-agent hiring at @hyperspell no resumes. no leetcode. you build an agent. our agent interviews yours if you can build a great agent to do the job, that's the proof you can do the job anyone can apply. we will interview every single agent

@AntaraWork4945 @conor_ai @hyperspell Email us at founders@ and we'll hook you up

@conor_ai @hyperspell @maebert few questions: 1. after creating it is there a form through which we have to submit? 2. any poc we can use to contact

@thesayannayak I created my first web pages raw-dogging notepad.exe in the 90s. Syntax highlighting is for the feeble minded.

Be honest, what age did you started coding?

Our in-house AI agent at @hyperspell, Cubert, is telling it like it is. I just wish I could follow any of it.