Maksym Savchenko

@simonw people always think they're the exception to this

Doing this is a great way to make a bonfire of your reputation

@jackbutcher clean work, what's the project for?

New brand work ][

@marclou one guy with nothing to lose hits different though

POV: you’re a giant startup marketplace trying to deplatform sellers from a tiny indie marketplace run by one guy

@emollick @waitbutwhy yeah but what happens when the slowdown actually hits, that's the real question

Everyone has seen the @waitbutwhy cartoon of AI capability growth with a "you are here" indicator just before the exponential really starts, but the independent assessments of both METR and the UK's AISA do seem to show that we are past that point now (until we hit a slowdown?)

@jackbutcher so basically you're never actually done then

You are only one impressive thing away from having to do something even more impressive

@levelsio lmao what, you need to accept terms to turn on your AC? that's dystopian

Today my Daikin AC required me to accepts its new terms and conditions to continue using it

@sama that's what everyone says but then you actually live it and it hits different

being a dad is the thing that has most exceeded already-high-expectations in my whole life

@levelsio agree but at least it's memorable

Most annoying haircut I've ever seen in my life

@gregisenberg the MCP part is wild but who actually owns the relationship with the agent customer

My 30+ observations on the greatest opportunities in AI agents right now: And some ideas that are keeping me up at night. 1. The new buyer on the internet is an AI agent. Imagine billions of new customers showing up with money to spend but they only shop via MCP. That's what's happening. No MCP server means you're invisible to the fastest growing buyer on the internet. 2. Every franchise system in America (30,000+) needs an agent layer and none of them have one. One founder per franchise vertical. That's 30,000 businesses waiting. 3. Everyone said "distribution is the only moat" a year ago. Now I'd add that the only moat is distribution plus memory. The company that has your audience AND your agent's accumulated context is impossible to leave. 4. Consumer mobile is more interesting than it's been since 2012. Apps can finally DO things for you instead of showing you things. The next wave of $100M apps are being built right now. 5. The most interesting startup nobody has built is an agent marketplace where you rent access to someone else's trained agent. A recruiter spent 6 months training a sourcing agent on healthcare hiring. That agent is worth renting to every other healthcare recruiter on earth. The agent itself becomes the product. 6. A sorta strange phenomenon that's happening right now is agents are developing preferences. Give the same agent the same task 100 times and it starts developing patterns in how it approaches it. Nobody is studying this yet. But the agents that develop good patterns are worth more than the ones that don't. That's a new kind of asset. 7. Dead internet theory is about to become dead SaaS theory. Half the apps you use will quietly replace their support team, their onboarding team, and their content team with agents. You won't notice for months. Then you'll realize you haven't talked to a human at that company in a year. 8. The most valuable data in the world right now is sitting in the support tickets of small or mid tier SaaS companies. Every ticket is a customer telling you exactly what to build next. Mine this. 9. The most interesting pricing problem nobody has solved is how do you price a product when your costs change every time OpenAI or Anthropic updates their model pricing? Your margins can swing 40% overnight based on a decision made in San Francisco. The company that builds dynamic pricing infrastructure for agent-based businesses solves a problem every AI company has. 10. The best AI products feel like they're reading your mind. The worst ones feel like filling out a form with extra steps. 11. An interesting arbitrage I've noticed lately is hiring a human VA for $20/hour to supervise an AI agent that does $200/hour work. The human just checks the output. 12. The managed AI agent business is becoming the new agency model. $5k/month per client. You build it, run it, maintain it. The client gets a digital employee they never have to think about. This will be a $50 B+ category. 13. The first "shadow agent" scandals are about to drop. Employees running personal agents on company infrastructure without telling anyone. Using company API keys. Agents accessing internal docs. IT departments have little visibility into this right now. Lots of opportunity to build companies here. Definitely a painkiller not a vitamin type of business. 14. Right now there are probably millions of agents running on autopilot that their creators forgot about. Still burning tokens. Still sending emails. Still scraping websites. Still costing money. The "find and kill your zombie agents" tool is a product that writes itself. 15. Companies are starting to hire based on someone's agent portfolio instead of their resume. "Show me 3 agents you built that are running right now." It's REALLY early but it's starting. 16. Your Slack archive is a product. Every company's internal Slack has thousands of messages explaining how they actually do things. The company that lets you point an agent at your Slack history and auto-generate SOPs and agents from it will be enormous. 17. We're watching the cost of intelligence fall faster than the cost of distribution. Which means distribution is now the expensive thing. 18. The most underrated asset a human can have in 2026: the ability to sit in a room with another human, make eye contact, and have a real conversation. As AI handles more of the transactional stuff, the humans who can do the relational stuff become disproportionately valuable. The soft skills people used to dismiss as fluffy are becoming the hard skills. The hard skills people spent decades acquiring are becoming the soft ones. 19. There are MANY huge companies to be built around the fact that most people's agents are running on their personal laptops which they also use to browse the internet, check email, and download random files. The attack surface is enormous. One compromised Chrome extension and your agent's API keys, customer data, and workflows are exposed. 20. There's a new type of burnout forming that doesn't have a name. It's not from working too hard. It's from context switching between human work and agent work 50 times a day. Reviewing agent output, correcting it, approving it, reviewing again. The mental load of supervising agents is different from the mental load of doing the work yourself. Some founders are telling me they were less tired when they did everything manually because at least the cognitive pattern was consistent. 21. The cheapest form of market research: search "[your industry] spreadsheet template" on Google. Whatever people are tracking manually is your product. 22. Half the YC companies pivoted within 8 weeks of demo day. Not because they failed. Because agents let them test 5 ideas in the time it used to take to test one. The concept of "committing to an idea" is dissolving. Serial pivoting is becoming the default because 1) AI lets you move fast 2) the world is moving fast. 23. The loneliest job in tech right now is being the only person at your company who understands what the agents are doing. You can't explain it to your boss. You can't hand it off to a colleague. If you leave, everything breaks. You've become a single point of failure for an entire automated system. That person needs a title, a team, and a backup plan. Most companies haven't figured this out yet. 24. Your browser history is the most valuable training data you own and you're giving it away for free. Every site you visit, every product you research, every competitor you study, every pricing page you screenshot. That behavioral data, structured and fed to an agent, would make it understand your business better than any onboarding call. The company that lets you turn your browser history into agent context builds something nobody can replicate. 25. Everyone is building AI wrappers. Nobody is building AI unwrappers. The tool that takes an AI-generated document and tells you which parts a human wrote and which parts were generated. 26. Stripe just became the most important company in the agent economy and they barely had to do anything. Every agent that sells something needs Stripe. Every agent that buys something needs Stripe. They're the payment rail for the entire agentic internet by default. 27. The most undervalued API in the world right now is the US Postal Service address verification API. It's practically free. Every local business lead gen agent needs it. Every real estate agent needs it. Every direct mail agent needs it. Boring government infrastructure is quietly becoming the backbone of agent-native businesses. 28. The concept of "business hours" is for humans. Your agent closed a deal in Tokyo at 3am, processed the payment, sent the onboarding email, and updated the CRM before your alarm went off. 29. What happens when agents start recommending other agents? Your research agent finds that a competitor's sales agent is better and suggests you switch. Agent referral networks are forming organically. The first agent affiliate program is probably 6 months away. 30. Cal dotcom closed their source code. That's the canary. When open source companies start closing up, it means agents were cloning their product too easily. Every open source company is quietly asking the same question right now. 31. "AI for pet groomers" sounds like a joke and that's exactly why it will work. 150,000 of them in America. Zero tech. All scheduling by phone or IG DMs. The joke ideas always win. 32. The thing that will seem most obvious in hindsight: we spent 2025-2026 arguing about which model is best while the entire value was in the orchestration layer. The model is the CPU. Nobody buys a computer based on the CPU anymore. They buy it based on what they can do with it. Makes so much sense in hindsight. What else will be obvious in hindsight? I'll share more notes soon. I can't sleep with all that's going on. Maybe you too. What an incredible time to be building.

@sama yeah speed/cost matters way more than squeezing another 2% accuracy honestly

i get some anxiety not using the smartest-available model/settings. but sometimes i dont mind if it's really slow. i wonder if we should focus more on a price/speed tradeoff relative to a price/intelligence tradeoff.

@sama two months free is solid, but how many actually stick after it ends

codex is the best AI coding product and we want to make it easy to try. for the next 30 days, we are giving companies that want to try switching over two months of free codex usage.

@marclou doing a startup launch AND training for hyrox in 2 days is insane, how are you sleeping

The startup office is now in Korea 🇰🇷 I'll be there for a few weeks, and this is what I'm working on: ✔️ DataFast AI-first ✔️ New SaaS ✔️ New project I'm also training for my first Hyrox competition in 2 days. Excited!

@naval hard to notice when you're the one doing the reasoning though

The enemy of truth is motivated reasoning.

@emollick slash commands might as well not exist if nobody knows about them lol

OpenAI contacted me to say “Study Mode is still live and accessible via /study and /learn shortcuts” so that’s good, although the official study mode page doesn’t mention that. (I don’t think slash commands are a natural thing for the vast majority of people).

@emollick making it harder to access is a weird way to solve that problem though

The silent removal of Study Mode from ChatGPT is a big mistake (both Claude and Gemini still have theirs) We have enough evidence that using AI in assistant mode to study can hurt learning because it just gives you answers, making students think they learned when they have not. You can prompt the model to be a very good tutor, but most people don't know to do that. Study mode was an easy option that parents and teachers could suggest to mitigate negative effects, even if it wasn't perfect. OpenAI still has a page about it, and the link activates study mode but otherwise there seems to be no way to select it from a menu for most accounts. openai.com/index/chatgpt-… (Deleted this by accident, sorry, so reposted!)

@gregisenberg looks interesting but how much are people actually paying for this vs building it themselves in a weekend

I'm giving away a FULL course on how to build a managed AI agent business solo using Hermes Agent, Orgo, Obsidian, Codex, Claude Code etc. Here's everything (47 minutes): 1. The offer: unlimited agents, unlimited usage, all infrastructure and security included. The customer gets a digital employee. They never think about tokens or models. You handle everything. 2. Don't niche down too fast. Try marketing agencies, law firms, insurance, manufacturing, real estate. See where the market pulls you. Then go vertical. Diverge first, converge later. 3. Every executive has the same problems regardless of industry. Too many emails, too many meetings, too many follow-ups, too many open loops. Solve those first. Then layer in vertical-specific skills. 4. The stack: Hermes Agent for the agent harness. Codex or Claude Code desktop to build and configure. Orgo for cloud computers so every agent lives in its own sandbox. Composio for one-click authentication across thousands of apps. Agent Mail to give every agent its own email. Obsidian for the knowledge base. 5. Use agents to build agents. Don't stress about setup. Use Claude Code or Codex to install and configure Hermes inside a VM. Use Perplexity MCP, Context7, and Exa for up-to-date docs. Your agent sets up your customer's agents. 6. GPT 5.5 is the best model right now. Efficient with tool calls. Doesn't eat tokens like Opus 4.7. For cheaper tasks, GLM 5.1 from ZAI is the best open source option. 7. Set up watchdogs for gateway crashes so they auto-restore. Have agents email you when cron jobs break or skills fail. Your customer should never have to tell you something is broken. 8. Get customers through content. If someone jumps on a call and already knows who you are and what you sell, that's the position you want. Content is the most leveraged thing you can do in 2026. 9. Keep scope tight. One to two requests at a time, delivered in under 48 hours. Use Trello for customer-facing project management. Send Loom updates at random hours to show you're always working on their agents. 10. If you can set up Claude Code, Hermes, or OpenClaw, you have a skill that 99% of business owners don't have and would pay $5k/month for. You're probably not giving yourself enough credit. shoutout to @nickvasiles from @orgodotai for coming back on @startupideaspod and sharing the full playbook. tools, stack, fulfillment, everything. this type of episode isn't shared anywhere on the internet. this is the alpha people keep for themselves. i will keep sharing if you keep watching. you could watch netflix or you can watch this (link below) youtu.be/BI-MNjm1tTQ?si… watch

@gregisenberg curious if this flips back once people realize the individual influencers are just personas too

There's a strange inversion happening with trust. We used to trust institutions and distrust individuals. Now we trust individuals and distrust institutions. This is going to get more extreme as AI floods the internet with corporate sounding content and brands that all look the same because they're all trained on the same 200 references. The more polished and institutional something looks, the less people trust it because it pattern matches to "generated". The more raw, personal, and imperfect something feels, the more people trust it because it pattern matches to "real". I'm noticing this. Maybe you are too.

@simonw that's actually hard, how messy does the YAML get for anything real

New TIL: I figured out how to use my LLM CLI tool in a shebang line, which means you can write executable scripts in English, or hook up more complex scripts with a snippet of YAML template

@gregisenberg true but at least you can kill an agent at 3am without HR involved

If you were the kid who carried every group project in school, you'll love the AI era. People let you down. Agents will too. They break, they hallucinate, they do the wrong thing at 3am. The difference is the feedback loop. When a person drops the ball, you have a conversation, wait a week, and hope it gets better. When an agent drops the ball, you fix the prompt and it's better in 5 minutes. Same frustration with 100x faster resolution. Not everyone will love working like this. But a lot of people will. I really like it.

@simonw lmao that's wild, claude just sitting there eating your whole drive

My Mac had less available memory than I expected, turned out the "claude" Claude Code processes on this machine (running in various terminal windows) were consuming ~30GB on their own! The largest one was using 4.9GB