TLDR Founders 2026-07-17

The Bay Area pulls almost 2.5x the capital of the entire New York ecosystem across every category combined. Value concentration in AI and B2B is more concentrated now than it was five years ago. Two metros combined take 72% of B2B dollars. Talent and capital didn't disperse - the narrative got ahead of the numbers.

Companies selling data to frontier AI labs know the cash is real, but the contracts are not forever. There are six categories of products being sold to frontier labs: hours, judgment, worlds, verdicts, bodies, and rights. Each generation of models graduates past the data that trained it, so labs need fresh data to keep pushing the frontier.

Agents at Replit now investigate production incidents, review pull requests, answer questions, analyze business data, triage support tickets, research sales accounts, and improve the systems that power the Replit Agent itself. The company's expanding system of agents takes goals from people, gathers context, performs work, checks the results, and escalates when human judgment is needed. People still decide on which problems matter, make difficult tradeoffs, exercise taste, and take responsibility for the outcome, but they increasingly do not perform every step required to get there.

The point of the onboarding process is to teach users how to use your product. The MCP onboarding process should give agents a step-by-step guide on how to walk users through their first time using your product. Agent-native support can be better than founder support. The field of agent-first interfaces is still in a nascent stage, so a lot can still change.

Many founders think they're running a SaaS company. However, if the team uses the platform to deliver work for clients who never log in, they're actually running a consulting company with good internal tools. This changes how founders should hire, price, and focus. Consulting can still be a good business, so either path can succeed. What doesn't work is drifting between being a consulting and a SaaS business.

Cerebras' employees ask the company's internal knowledge base more than 15,000 questions every day. It has become one of the company's most widely adopted internal tools since launching 3 months ago. The knowledge base is used by humans, automations, and agents. This post details how Cerebras Knowledge was built.

Someone can discover you in ChatGPT, Google your company a week later, and show up in GA4 as Direct traffic. That is why AI search can look tiny in your analytics even when customers are using it. Only about 1% of people click citations in Google AI Overviews, and 70.6% of AI-referred visits are recorded as Direct. Before deciding AI search is doing nothing, add ChatGPT, Claude, Perplexity, and Google AI to your "How did you hear about us?" question. Then tag the same answers in your sales calls.

1Password for Claude is built on a zero-exposure architecture that only grants access at runtime. Claude never sees the vault item, password, or one-time code, and access is scoped to the current task and ends when the task is complete. Access requires user-consented biometric approval, and credentials are injected directly into the page. 1Password for Claude is now available for Mac for business, family, and individual plans.

Muse Spark 1.1 is now available on OpenRouter for developers in the US. Muse Spark is a multimodal reasoning model from Meta. Built for agentic tasks, it accepts text, images, video, audio, and PDF documents and returns text. The model has a 1-million-token context window. It is designed to orchestrate multi-agent workflows and supports structured output, parallel function calling, built-in search with citations, and configurable reasoning effort.

Natural language is a far more useful form of input than the indirect signals the industry had to live with over the past 20 years. It allows companies to read exactly what users want to accomplish and whether or not they're satisfied. Natural language inputs, paired with other signals, can help companies see the relationship between the performance of agents and revenue. Agent observability is not only an engineering concern - CEOs, product leaders, and customer-facing teams should be reading this data too.

It's an open secret that AI agents aren't reliable because they're stochastic, prone to lying, expensive, and make mistakes. Businesses are responding by selling better cages, but AI is not deterministic. Nobody survived the cloud wars by having a better shovel - they survived by building moats. These companies are competing with model improvements, and the models will eventually win.

Venture firms can no longer rely on IPOs, M&A, or secondary buyers to convert portfolio markups into DPI. Healthy private companies may need cash-flow-driven buybacks and capital return programs, supported by duration-matched credit, to create liquidity without sacrificing growth.

Open source AI is funded by companies shipping a deliberately unremarkable, general-purpose base cheaply and letting customers build value through customization.

Product executives, especially those skilled in building with modern tools, are now commanding salaries akin to pro-athletes due to high demand and scarcity.

If a customer says "interesting" after your demo, they are probably about to ghost you.

Investors, LLMs, and probably your friends all say "taste" is the only thing left to defend now that AI can write the code.