35,595 results · Most relevant
Turning everything over to computers with no humans in the loop causes systems to go off the rails when something weird happens; keep humans in the loop for AI systems, including the AI admin idea.
Before deciding on GPU count or deployment, ask what model quality and what tokens-per-second speed you actually want. Customers typically target 20-30 tokens per second (ChatGPT speed) up to 60, depending on how long the chain of tasks is.
Split API keys by department and by major use case so you know where all your token spend is. Know the token spend per department and per team so you can start seeing which team is performing better and what works. Check which teams spend the most tokens and whether they are productive. Use per-team token usage data to advise on dedicated capacity purchases, whether for open or closed models.
The speaker is not expecting the audience to start building their own custom harnesses from scratch because great solutions already exist in the market (e.g., Cloud Code, Codex). But you can build custom harnesses for specific problems without them being as sophisticated as creating a new Cloud Code. Simplify the problems you are trying to solve and then build custom harnesses that attach onto that, rather than running all your work in one sophisticated custom harness. Consultants are a perfect use case: they can create one that works for a financial agent and then adapt it for different customers, data types, or constraints.
creator's guidance; consultant case is hypothetical
Oliver Kharraz describes how ZocDoc still operates its destination-website model for finding and booking doctors, but has evolved into an infrastructure layer that payers, AI platforms, and consumer companies can plug into for seamless appointment scheduling.
CJ Avilla explains the orchestration model: a lead agent creates a plan, automatically spins sub-agents up and down based on its own capability, and accomplishes discrete verifiable units of work across phases without the user managing interconnections.
Storing API keys as plain text on disk (even hidden files or secret directories) risks the credential leaking, whether via a user prompting Claude or your product to extract it, or through prompt injection or user manipulation. Putting an API key into a markdown file given to Claude is definitely bad — do not do that.
Anthropic created an entire suite of abstractions to ensure credentials never have to touch the sandbox — an incredibly important decoupling. Every container that has access to any secret uses totally fake secrets that don't actually work if used to call the real API; credentials are substituted for real values only at the egress proxy layer on the way out of the container. So there is no way physically to prompt Claude or access the sandbox and get the credential off the sandbox or off Claude's brain. Take credential safety seriously — by default you should not be able to foot-gun yourself into leaking credentials when using Claude Managed Agents.
described
When a task has too many pieces for one agent, dynamic workflows let the agent write the workflow itself and a server runs it in the background.
A new code layer can be created directly on the Figma canvas, with presets on the right panel for starting from particular sizes, and work begins by prompting directly into the layer.
A code layer can be started from existing code by uploading a local folder or cloning a GitHub repository (with a connected, authorized GitHub account), keeping the repo on the main branch.
Hitting the preview button on a code layer shows the app running live on the canvas; a cloned repository becomes a fully live, previewable app (e.g., a marketplace browsing experience plus a map view).
The extract designs feature pulls code into Figma as actual editable design layers: extract the current state of a code layer, select an individual element to extract just that element, or record a flow (navigate pages, then stop) to capture multiple screens as full layers.
Moving to a 3PL was probably the best move the brand made; self-fulfillment is fun until it isn't, and the decision ties back to being present with your kids rather than living in a warehouse 12 hours a day.
The guest offers the theory that going product-first turns you into a product-led brand instead of a community-led one. Dad Gang built community by being themselves rather than being strategic, running a private Facebook group of 15,000 dads who interact with them and give product input, and talking with community members daily in DMs, emails, and comments instead of automating engagement.
partial
Build the audience first before expanding the product line; starting product-first risks becoming a product-led brand rather than a community-led brand, and if you have an audience to listen to, build it first before guessing what new colors to release.
Keep overhead low by outsourcing functions like 3PL and customer service; a lean structure also lets founders go full-time sooner.
The founders advise focusing on getting really good at selling one product (their black-and-white trucker hat) before adding other colors or SKUs, because there's a timing for everything; adding many SKUs early is a massive headache for apparel brands, and outside noise will push you to complicate your life with new colors, shapes, t-shirts, and sweatpants.
The speaker argues that moving to a 3PL was probably the best move they made because they could scale alongside a partner growing at the same time, and recommends choosing a 3PL that is also a growing startup so you can go through growing pains together.
The speaker asserts that the biggest signal for building a high-performance team is who you fire — even bigger than promotions. He immediately qualifies this: he is not advocating a fire-everybody culture. The point is to enforce a minimum bar, not to treat termination as the default tool.
Choose sources → to get a library of your own and a daily brief.