Subscribe for startup ideas & tutorials on how to use AI to build & grow your startup. Hey, I’m Greg - CEO of Late Checkout and ex-advisor to Reddit, TikTok, and more. I post videos here from my podcast top 0.1% tech podcast called The Startup Ideas Podcast. You can also listen to The Startup Ideas Podcast on Apple Podcasts and Spotify. I keep it practical and never gatekeep the sauce. I share what I'm learning in real-time. Happy building
In Hyper Agent's early adoption base, small shops have become surprisingly sophisticated very quickly and are running their operations in genuinely game-changing ways. The reason is simple: a 50,000-person company can't pivot on a dime — th
The design philosophy behind Hyper Agent is to serve the full range of users — from power users who want granular control over how the agent is set up, so they can squeeze out maximum bleeding-edge performance, all the way to people who jus
Greg Isenberg frames platform selection as an investment decision: when you pick a platform — whether it's OpenClaw, Hyper Agent, Codex, or something else — you are investing in an ecosystem. That framing matters because the switching costs
MembersAs users accumulate memories across many agents, managing those memories becomes its own problem. Hyper Agent recently shipped a defrag tool to address this. It auto-suggests related memories clustered by both keyword and embedding similari
Howie draws an explicit parallel to Apple and personal computing. Frontier agents — across every platform — are all going to get better, smarter, and faster. Hyper Agent isn't going to be the only powerful agent product. But just as Apple u
MembersHyperagent recently shipped a new onboarding flow that moves beyond a generic empty canvas. When you first land in, it suggests connecting your Gmail, Slack, Notion, and meeting tools like Granola. It then goes and reads through your recent
One of the most powerful things Hyperagent can do is handle APIs it doesn't already have a connector for. You simply ask it to build a custom skill for a given API — say, Twilio — and it will go research the API documentation on its own, wr
MembersThere's a subset of emails that are low-stakes enough that you'd want an AI agent to not only draft a reply but, if it feels confident the situation isn't sensitive, just go ahead and respond automatically. Simple examples include inbound m
You can turn any thread or task in Hyper Agent into a recurring scheduled job simply by telling it in the thread: 'Can you turn this into a recurring daily email at 8 a.m.?' The agent will save the thread as an agent with a run schedule — e
When you get feedback on an AI-generated output — say a tweet that sounds too formal — you have several options right away. You can update the underlying skill based on that feedback, have the agent immediately regenerate a new draft that s
MembersWhen reviewing AI-generated tweet drafts, Greg Isenberg noted a common quality problem: the output felt too formal and stiff rather than sounding friend-to-friend, which is the tone that tends to perform well on social media. The fix isn't
MembersOne limitation of human-written tweets, as Greg Isenberg acknowledges, is that they're typically just hot takes with no supporting data. What becomes possible with a team of AI hyper-agents doing research in the background is that those tak
The biggest mistake people make when learning to build with AI agents is trying too sporadically — picking it up occasionally, then dropping it. Greg Isenberg recommends committing to a set period — 30, 60, or 90 days — with a literal calen
Getting great outputs from AI is like learning tennis — when you first start, you're bad at it, and you almost don't want to show up because you feel incompetent. But you go through the messy middle, you get better and better, and eventuall
A lot of people one-shot something with an AI agent, it's not quite as profound as what they hoped for, and they give up. But agents are powerful enough to do almost anything you want them to do — the issue is never whether they're capable.
MembersHyper Agent has a feature called rubrics — essentially an eval rubric that lets you define what 'good' looks like for a specific type of task. For example, you could create a rubric called 'What makes great Greg Isenberg content?' and pin i
MembersAgents in Hyper Agent don't just accumulate new memories over time — they actively participate in their own improvement. The platform has an automatic self-improvement loop where an agent can suggest to you, 'Hey, maybe you should add this
Unlike Manus or Perplexity Computer — where the nominal experience is using one awesome agent directly — Hyper Agent is designed from day one with scalability and deployability in mind. Once you have an agent that works for you, you can dep
Against OpenClaude, the distinction is fairly straightforward. Hyper Agent is a more general-purpose agent platform that is turnkey, ready to go, safe and secure by default, and cloud-native. OpenClaude, by contrast, feels quite raw — when
Perplexity Computer and Manus are the closest competitive comparisons to Hyper Agent. Manus was truly groundbreaking when it first launched — it was the first real 'holy crap, yolo agent' that showed what autonomous agents could do, even be
MembersSkills are described as the single most important concept in the frontier-agent world. The underlying models are already generally intelligent — think of it like hiring an Albert Einstein who is brilliant in the abstract but doesn't know re
Any agent you build in Hyperagent can be pushed into Slack with a single click. The agent configuration lets you choose a channel, set the trigger mode (mention-only or listen to the full channel), and toggle auto-responses. In 'listen to c
Hyperagent ships with powerful tools out of the box, including Google Maps integration. One example: you can instruct it to find real Street View imagery of billboard locations — it knows how to locate actual points of interest, pull the im
MembersBecause Hyper Agent has full coding capability, it doesn't stop at producing a research brief — it can go ahead and build a v1 of the product it just researched. In this demo, it created a working app called Blockpulse: a clean, interactive
Hyper Agent runs a full computer in the cloud — not on your own machine — with a full sandbox environment and complete coding capability. You can connect it to all your accounts: Slack, Granola, email, and more. It can send things on your b
Hyper Agent is built on the same design philosophy and obsession with great UX that the Airtable team applied to the no-code app category ten years ago, now brought to agents. Just as Airtable distilled the technical complexity of a Rails a
MembersHyper Agent is built around the idea that most other agent products out there — think Open Cloud and similar — are more like Linux: powerful but raw. Hyper Agent is the Mac version of that. It's a multi-window canvas workspace where an auto
There's a parallel worth noticing between humanoid robots and AI agents. Robots are converging on a humanoid form factor not because it's philosophically ideal, but because all of our existing infrastructure — homes, construction sites, fac
The future of work, in one image, is an Agent Command Center — a single operator overseeing a fleet of agents handling Customer Intel, Content Production, Competitive Research, Lead Enrichment, Inbox Triage, M&A Diligence, and more. A termi
As a concrete illustration of AI-agent value, the speaker described a recent board memo he sent to his full board and major investor list. Much of the memo was researched and crafted by their AI agent (Hyper Agent), informed by his own inst
MembersA lot of people complain about the cost per token of frontier models — Opus 4.5 and GPT-5 are expensive, and even open-source alternatives aren't free. But this objection misses the point. In the old days of software, a lot of things were c
MembersThe unit economics comparison between human agents and AI agents is dramatic. Cost per customer interaction drops from $3–$6 for a human to $0.25–$0.50 for an AI agent — an 88% reduction. First response time falls from over 6 hours to under
Using is believing. It is genuinely hard to fully grasp what frontier agents can do if you haven't gone hands-on and spent at least a full weekend with them — and that means more than a superficial, naive one-shot prompt like asking who wil
A trillion dollars is actually an underestimate. The real total addressable market for AI agents is essentially all white-collar labor — many tens of trillions of dollars, even in the western hemisphere alone. Once agents can do nearly anyt
Even in a basic synchronous chat interaction — not a full autonomous agent loop — you can give these frontier models the hardest, meatiest problems in domains like management consulting and receive genuinely expert-level answers. The model
Agents reached a genuine breakthrough roughly four to five months ago. Developers felt this most clearly with Opus 4.5, which set a new high-water mark: for the first time, the model could take on a task that would have taken a real human e
Sequoia has laid out a chart showing roughly a trillion dollars of addressable market within AI agents, spanning industries like consulting ($300B+), IT managed services ($100–200B), accounting ($50–80B), and freight ($100B+). The framing i
Howie Liu's new product, Hyperagent.com, is an AI agent builder that lets users create digital employees and build apps around different business ideas. To support the Startup Ideas podcast community, Howie is committing one million dollars
Sequoia has published a two-by-two matrix mapping agentic software opportunities across two axes: Copilot (human-assisted) versus Autopilot (outsourced), and Match (in-housed) versus Next Wave. The matrix maps this against Judgment versus I
MembersRather than using a traditional IDE, Howie Liu describes his own development workflow on Hyper Agent as running approximately 30 different Claude Code instances in parallel. Each instance is coupled to a browser and operates fully autonomou
MembersHowie Liu argues that the ~50% share of AI agent tool calls attributed to software engineering is actually an overestimate of how many developers are truly working at the frontier. Most engineers are still using AI in a 'gen one' mode — aut
Howie Liu's reaction to Sequoia's agent deployment chart is that it absolutely reflects the underpenetration of AI in industries that could already be disrupted or benefit with even today's AI capabilities. His view: if you took frontier ag
According to a Sequoia chart on AI agent deployment by domain, software engineering accounts for nearly 50% of all tool calls made by agents. The next largest categories are back-office automation at 9.1%, marketing and copywriting at 4.4%,
Every piece here is something Greg Isenberg can talk you through, in their own words.