RecommendationThe speakers recommend giving Claude access to GitHub repositories so they mount inside the container where Claude operates, uploading a pre-computed dependency graph as a file and pointing Claude to its path instead of regenerating it, maintaining a crisp Bazel build with a clean dependency tree, and remounting the repository on the container for every PR review like a CI job checkout. If the codebase lacks an explicit dependency mapping, Claude will have to search files for symbols to build the graph.
The symbol-search behavior is described with hedging ('probably').
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
RecommendationThe speakers describe running hill climbing human-driven — generate two outputs, say which one you prefer, and it generates two more, iterating on its own prompt. In Claude Code you can run '/claude API hill climb' and tell it what you want to iteratively improve on; taking an off-the-shelf watch-list-scan prompt and saying 'Claude API hill climb this prompt' grinds until it produces a prompt that meets your needs. They also note that without the hill climb skill or dreaming, the flow won't self-improve by seeing what it did in the past.
The charades-prompt frame-rate detail is anecdotal.
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
RecommendationYou can build the explainer-video flow yourself in Claude Code with the Cloud API skill (manage agent onboard, explainer, video maker); just ask a question like 'What is DNS? How does DNS even work?' and it will generate a flow.
endorsed
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
RecommendationJed recommends spinning up a reviewer agent with a prompt listing what to look for, pulling session transcripts via API, and asking whether agents did anything unexpected, accessed anything they weren't supposed to, or called odd tools — reviewing the past week of transcripts.
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
RecommendationUse MCP tunnels to give agents private MCP access to data behind a VPN without exposing anything over the open internet, using Cloudflare tunnels; the connection runs directly from wherever your private MCP server is running to the place the MCP calls happen. Connecting over the open internet is the closest alternative approach, and running a private MCP server with a direct connection is another option.
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
RecommendationLimit the network access of your agent physically using the environment configuration: restrict the environment to only specific domains (e.g. only api.twilio.com) so all other egress fails at the firewall layer, keeping API keys hidden away from potential theft.
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
Recommendation · MoneyJason recommends that if you want to build your own type of 10K or your own headless instance, try it three ways — Claudeforce, Replit, and Lovable — to see what today's connectors can actually do. Going headless unlocks integrations: once the team went headless, they could hook up QuickBooks, Bill, AgentForce, and other applications. Their own 10K agent, built on Replit, has run stably and securely in production for a year connected to 20+ different apps.
None; direct recommendation backed by their experience.
The Official SaaStr Podcast: SaaS | Founders | Investors · SaaStr 882: Claudeforce vs. Replit vs. Lovable. Who wins? (The Agents #016)
Recommendation · MoneyThe founders emphasize making content nonstop, documenting everything, and posting on Instagram every single day without missing a day.
endorsed
Foundr · We Started Dad Gang with $750 and One Hat - and did $40M in 4 Years
Recommendation · MoneyThe founders recommend outsourcing 3PL and customer service to keep overhead really low, and calculating whether a hire is truly needed or whether the founding team can get it done themselves.
endorsed
Foundr · We Started Dad Gang with $750 and One Hat - and did $40M in 4 Years
Recommendation · MoneyThe founders recommend launching a product by posting it on personal accounts (Instagram, Facebook, TikTok) accompanied by the story of why it was made, which they say connected with people.
endorsed
Foundr · We Started Dad Gang with $750 and One Hat - and did $40M in 4 Years
Recommendation · MoneyThe speaker recommends filming and posting time-lapse videos of packing orders as content while fulfilling from home; he was posting packing videos nonstop, and the founders made great content filming Bart and his wife packaging orders out of the garage.
Foundr · We Started Dad Gang with $750 and One Hat - and did $40M in 4 Years
Recommendation · MoneyThe founders recommend continuing to post content and tell the brand story during stockouts (for 30-40 days) rather than going silent, using scarcity as a marketing asset through authenticity about fatherhood.
endorsed
Foundr · We Started Dad Gang with $750 and One Hat - and did $40M in 4 Years
Recommendation · BodyThe host advises that exercises like Zurcher squats, Anderson squats, and back squats can't be randomly thrown into a program if you want to get stupid strong: place heavy strength exercises roughly second in the session, progress them week over week with a deliberate plan, and perform squats explosively to train recruitment speed. He gives an example of paused back squats for four sets of three, building up over a couple sets, then one drop set of five.
Garage Strength · 5 Old-School Exercises That Get You Stupid Strong
Recommendation · BodyDo four sets of 17 reps on the hack lift, loading it with 70 to 120 kilos (155 to 265 pounds) and hammering the reps. Place the exercise a little bit later in the workout, like a hypertrophy or accessory movement, and do it consistently at least once a week. The lift is absolutely fantastic, especially if you have smaller quads.
Garage Strength · 5 Old-School Exercises That Get You Stupid Strong
RecommendationWhen using DEL data for ML training, condition models on the existence of the DNA barcode rather than trying to remove its influence — the binding measurements are real physical signal. In practice, encode the barcode as a placeholder marking its position on the molecule.
Latent.Space · Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio
RecommendationFor designing pocket-breaking mutations, Quigley points to AlphaFold, Boltz, and similar structure models. For poorly characterized proteins, Leash combines evolutionary conservation analysis with a staff structural biologist's judgment to pick which amino acids to mutate.
attributed to Leash's practice
Latent.Space · Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio
RecommendationTest the bead alone and check for pan-protein stickiness signals so you can make a reasonable comparison to the actual binding measurement — you have to think hard about these controls.
Latent.Space · Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio
RecommendationSplit training and test sets by building-block clusters so no entire cluster overlaps between test and train. If you split by shared building blocks instead, performance is inflated by memorization (models memorize individual building blocks, implying similar building blocks behave similarly), so cluster-based splits are needed to reveal true generalization.
Latent.Space · Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio
Recommendation · MoneyRandy Hunt recommends that a specialist's first job should be building an agent that scales their own expertise and best practices — for example, a solo UX writer encoding their best practices so the whole company benefits while the specialist monitors and tunes it. He reports Notion's UX writing agent became people's favorite tool company-wide within roughly two months, and that colleagues also learned what the specialist does and when to reach out directly.
Design Better · Randy Hunt: Notion and Etsy design leader on why the best design teams stay small
Recommendation · MoneyRandy Hunt recommends using AI research agents with access to historical research context to answer 'have we ever talked to customers about this?' in real time during the design process, and querying them for prior customer interviews and past findings on a topic before starting fresh. He discourages relying on the old workflow of asking the research team in Slack, waiting on a person who may have left, or an unfinished research repository.
Also discussed by Karri Saarinen on Lenny's Podcast (unclassified source — not counted until classified). Not counted as independent sources.
Design Better · Randy Hunt: Notion and Etsy design leader on why the best design teams stay small