Principle · MoneyTurn-based auto-routing needs to be locked down to at most two to three models. If you put in 30 models and it keeps changing the model every turn, the low cache hit rate will wreck your optimization. Two models tend to be okay. Task and organization make the most sense for routing; request-level routing didn't make sense in the speaker's experience.
AI Engineer · The AI Bill Is Too Damn High: Tokenomics 101 — Eugene Cheah, Featherless AI
Report · MoneyThe agent loop controls everything that it connects to. In context assembly, there is a maximum number of defined tools or skills to retrieve — the default number is three. Every time a new request comes in, the model itself determines what is the thing to do, then assembles the context with the things it needs to perform, including retrieving the three most relevant tools and skills.
described as the workshop's mechanism; default of three is the speaker's recollection
AI Engineer · Memory-Aware Agent Harnesses for Anti-Money Laundering — Nacho Martinez, Oracle
Report · MoneyOliver Kharraz reports that the average wait time to see a primary care doctor nationwide was roughly 17 days when Zocdoc started and is now 31 days — it has gotten worse. The interviewer separately recalls an average wait of about 45 days for a primary care appointment in Boston around the time he left Kairos. By contrast, using Zocdoc, a patient can typically see a doctor within three days or less (72 hours) across a broad range of specialties.
The a16z Show · Why AI Could Finally Fix Healthcare Access | Zocdoc CEO Oliver Kharraz
Creator perspectiveYou can prompt Claude to give you reports or update a certain report as the run goes, typically delivered as markdown files you can open. If Claude writes to its outputs directory, you get a set of durable downloadable files via the v1 files API service, regardless of whether you're using flows. Run outputs can also be inspected as commits to GitHub, as in CJ's operating-system run.
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
Creator perspectiveYou can interrupt a running flow session to stop it and redirect it, e.g., 'you're going the wrong direction, please stop right here.' If you find new information mid-run that changes the trajectory, don't burn two days of tokens and start from scratch — instead, interrupt the session and provide the new information. When you interrupt, the primary agent has the full state of where the flow is in its process and can inspect where it left off to kick off a new flow that accounts for both the new information and the past history.
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
Creator perspectiveTo run agents in a Mac environment or other non-hosted environments, use the custom tools abstraction, which is a way for the model to kick out to you to ask you to take some action. Define a custom tool for your agent as part of the workflow run. When a custom tool call fires for running something on a local Mac OS machine or another arbitrary computer, do whatever you want on that machine and then return the result back to the model.
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
ReportWithin the Claude Managed Agents product surface shipped in April, Anthropic announced on Friday a new feature called Flows — essentially the next version of multi-agent orchestration where Claude can fully orchestrate the entirety of its own multi-agent topology. Anthropic envisions a future where Claude orchestrates itself, mapping interconnected webs to real-world organizations, with Flows as the way to do that on their API surfaces.
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
Creator perspectiveTo allow human feedback as part of a flow, implement it with custom tools: give Claude a 'request human feedback' tool, and Claude will request feedback whenever it needs it, with you providing whatever response you want from that tool call.
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
Creator perspectivePass pre-configured sub-agents in the agent config if you have agents that specialize in certain activities important to your research process. Set up individual agents that specialize in specific data sources like Yahoo Finance, news sources, or the SEC's Edgar website. Give agents access to custom skills, MCP servers, and your own custom tools that run inside the environment. You can also run the agent environment self-hosted in your own infrastructure by bringing your own sandboxes.
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
Creator perspectiveTo create a multi-agent flow, pass a multi-agent config when creating the agent. To get a flow up and running, send a first event to the lead agent saying 'run a workflow'. To enable flows, just tell Claude to use multi-agent and to run a workflow.
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
Creator perspectiveUse Haiku for the scan phase because it's a fast, efficient model that just goes out and scans the internet — spin up two Haiku agents per ticker for a comprehensive scan as an affordable phase. Invest more in the review phase when details are being missed, e.g., by using a Fable agent to review all the details in depth, possibly two or three agents per ticker at review.
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
Creator perspectiveTo control file-system access per agent, use the tool permissions abstraction: when an agent requests to read a file, check the request and the agent making it, and deny it if that agent should not have access. To restrict agents to a specific subset of directory paths, prompt them to focus only on those paths — there is no permissioning layer for directory paths. A permissioning layer for directory paths would be a useful feature, and the team may take the idea back and build something better.
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
Creator perspectiveDrive Claude managed agents and generally use the Claude API from inside Claude Code. Direct Claude Code to spin up resources (e.g., a managed agent that does X) and it will use the ant CLI to drive and interact with the Claude API. Instead of using the console, use Claude Code and say: spin up a managed agent that does X. If you don't want to use the console, jumping in via Claude Code is a recommended way to experiment with managed agents.
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
Creator perspective · MoneyWhen asked about AEO tooling, the 10K agent said there is no single AEO platform it would recommend yet, and suggested building an AEO dashboard in-house instead of buying Profound. AEO/GEO tools like Profound and Peak are good for benchmarking but historically weak at recommending actions — an agent can build its own visibility tracking and act on it. A free trial of Peak AI can run in parallel to benchmark against the in-house dashboard.
The speaker notes they haven't personally tried Profound or Peak.
The Official SaaStr Podcast: SaaS | Founders | Investors · SaaStr 882: Claudeforce vs. Replit vs. Lovable. Who wins? (The Agents #016)
Creator perspective · MoneyReacting to Lovable's unique recommendation, the speaker endorses rescuing deals past their close date or forecast close date first to keep the pipeline honest, agreeing it is probably the number one thing to do intuitively.
Framed as the speaker's agreement/intuition.
The Official SaaStr Podcast: SaaS | Founders | Investors · SaaStr 882: Claudeforce vs. Replit vs. Lovable. Who wins? (The Agents #016)
ReportThe host reports that smart TVs continue uploading data even when all smart features are turned off and the TV is used purely as an HDMI input, and that a Samsung TV transferred about one kilobyte per minute while playing content off an Apple TV.
Slidebean · The Most Profitable Part of Your TV
ReportSmart TVs upload data even when used as a dumb HDMI display with all smart features turned off. Samsung TVs send approximately one kilobyte per minute of data even when playing content via Apple TV over HDMI. The creators originally shot tutorials on disabling TV tracking settings on each screen, but question whether tutorials still stop the data collection now. The creators are not sure about other TV brands' data behavior. These are Latin American TVs whose terms of service are likely looser than those used in the US.
Slidebean · The Most Profitable Part of Your TV
Report · MoneyDad Gang operates as an eight-figure annual business with zero full-time employees, a few contractors, and an outsourced 3PL.
Foundr · We Started Dad Gang with $750 and One Hat - and did $40M in 4 Years
Report · MoneyVaynerSports reached out saying Gary loved what they were building and there were a ton of synergies, leading to Gary getting involved; the founders say the connection came simply from working with his team for a long time and loving what he built. Gary is an equity partner, not just an investor, calls himself their 'bat phone', and they've been working with him for only a few months. The founders compare it to LeBron wanting to play on your basketball team.
Foundr · We Started Dad Gang with $750 and One Hat - and did $40M in 4 Years
Principle · MoneyThe founders use AI to run reports, build spreadsheets and formulas that used to take all day (now 20 minutes), build website elements and design emails with tools like Manus and Claude, turn design ideas into outputs without Photoshop or Illustrator training, and design high-converting pop-out slide carts without a developer. They describe AI as an accelerator that lets them get far more done independently — work they might otherwise have hired five people for — and say it means they don't need to hire operational administration staff anytime soon.
Foundr · We Started Dad Gang with $750 and One Hat - and did $40M in 4 Years