Set hard rate limits and budgets that halt, not just alert
Andrey Rybka recommends setting hard rate limits and budgets on agent actions, plus extra guardrails, and making sure there are budgets that halt rather than just alert.
40,293 results · Most relevant
Andrey Rybka recommends setting hard rate limits and budgets on agent actions, plus extra guardrails, and making sure there are budgets that halt rather than just alert.
The speaker recommends starting agents fresh with no memories rather than persisting memory across sessions, and, where memory is used, checking it for malicious instructions because hidden instructions can persist as sleeper-agent behavior.
The speaker recommends not ingesting every package on day zero: delay ingestion by a week or two, verify the supply chain, and pair this with a deterministic invocation broker, least-privileged tools, and egress controls.
Exchanges based on the GPU hour (futures and derivatives) are coming, and market participants can use them to hedge out balance sheet risk and get into more interesting financial agreements. People are utilizing futures and are interested in utilizing futures and other options to hedge out the risk they currently have on the books; there is demand for this.
Farza describes his own process for launch videos: he writes every line of the script days before recording, then rehearses each line repeatedly in his head in the shower or on the way to work, asking how to say each line. He also visualizes himself delivering each line about 10 times before getting in front of the camera, so by the time he films he has already visualized it many times.
Number one rule for any platform (Twitter, Instagram, YouTube): would you stop on this? Would you even read the first sentence? If you wouldn't read the first sentence, there's no way you're going into it.
To find more customers like your natural adopters, ask what channels those specific people (e.g., design agency founders) hang out on — the answer is probably yes, they're on TikTok/Instagram/YouTube. Good default ways to find design agency customers: look on LinkedIn for people who list design agencies as their workplace, go to communities they hang out in, or visit specialized design websites. And make TikToks targeted at that specific customer segment as a lead-generation channel.
Farza recommends that if you're on your 15th, 20th, or 30th video, you try writing for just 10 minutes about what the video should be and then record it. He explains that the more you think through a video before recording, the better it can possibly get, because you can catch a bad idea before filming.
GaryVee advises that building something on the side while working a full-time job should start with passion: you cannot muster the energy to work from 7pm to midnight after a long day unless it's something you actually like.
Andy Jassy observes that most startups today start agentically and AI-native from the start, and predicts larger established companies will have to figure out how to build at that same agentic pace — some will, while those that move slowly will be more vulnerable to startups.
prediction, not established fact
Carmen Li recommends performing KYB/KYC verification on clients before extending compute credit or reserve contracts, to ensure they can pay at a scheduled rate.
When adopting outbound agents, pick one tool, run a small test, get it working even for just a subset of your outbound, push it as far as you can, then add another when you see gaps — and don't buy every tool the speakers use, since the stack is in flux and converging.
Consequential creator advice requires independent verification before Today can promote it.
Start with one surface (e.g., inbound only for the first week) and add more integrations over weeks rather than hooking up all tools at once.
When an inbound lead arrives, the prospector first surfaces all of the lead's data, then researches the pitch by searching first-party data first, then looks at external signals (Exa for company data, Sumble for company hierarchy and stakeholder email enrichment, CoreSignal for the person's LinkedIn).
If your site lacks a robots.txt or LLMs.txt, ask your agent to help create one and add it to your website so agents can read it and recommend/surface you — LLM-facing docs are a distinct content type written for agents to read, not for humans.
Give each agent a specific goal per channel (inbound → book a meeting, questions → escalate to a human, outbound → book a meeting or buy tickets) so it works relentlessly until the goal is reached, and don't let it attempt to answer everything.
When an inbound lead arrives, have the agent first check first-party internal data (Salesforce, spreadsheets, prior contact history) before responding, so responses carry historical account context instead of being blind.
Jarmak describes running regular audits in her software factory with approval and required checks, automating as much work as possible while not trusting the systems.
For lower-risk decisions, Jarmak has agents identify what the next action would be, why a human needs to weigh in, what the operating scope is, and potential suggested human checks — her own setup in her software factory.
Anekha Sokhal recommends that when handling customer data you avoid sending it to third-party processors if you can build in-house — Moshi built a vision language model to understand clients' messy PDF data, taking uploaded files, breaking them down, and rebuilding them in its own database (some data types require connecting directly to client systems instead). If you don't have a team of engineers, off-the-shelf document processing solutions are a great option; if you do have engineers, they will likely want to build the models. If you do use a third party, ensure you have agreement on how they process the data and what they do with it.
Endorsed guidance conditioned on team capability; no measured outcomes cited.
Choose sources → to get a library of your own and a daily brief.