TLDR Founders 2026-06-03

Most agents run on borrowed sessions. The user authenticates, the agent inherits their full OAuth token, and a support agent that looks up invoices can also refund, delete, and modify. The agent needed read access to one record. The blast radius is every account in the system.WorkOS is built to solve it. Agents need their own identity, scoped credentials, and permission boundaries. WorkOS ships the OAuth flows, RBAC, and fine-grained authorization to scope what each agent can do, plus audit logs tied to every agent, every action, every authorization.Read the full guide →

AI token costs are rising dramatically, leading companies to reconsider or cancel AI initiatives. EntelligenceAI reports that only 18% of AI spending results in productive output, causing concerns about the sustainability of current AI practices. Uber and others highlight the difficulty in justifying excess spending without direct productivity gains, signaling possible slowed growth in the sector.

There's a quiet land grab happening in AI dev tools right now. Everyone wants to be the platform, but the market decides who actually becomes one, and that's whoever you can't leave without it feeling like surgery. The way you get there is to own a "primitive": one small, painful, unavoidable job that everything else ends up depending on, the way Stripe owns payments or Twilio owns texts.

37signals runs almost entirely on writing instead of meetings, and Jason Fried just published the full guide to how. The bet is that long-form, written, async communication beats talking for nearly everything that matters. You can feel the philosophy in the rules of thumb: meetings are a last resort, writing solidifies while chat dissolves, a one-hour meeting with five people is really a five-hour meeting, and ASAP is poison.

Mitchell from Shown Media argues the invisible pattern behind Claude, OpenAI, Google, and Higgsfield is not researcher pay or fundraising but a release cadence that turns every ship into an event, executed through a system most companies do not have. Step one is the viral launch, and step two flips the typical model where marketing receives finished features and ships a blog post: top companies give marketing a seat at the monthly roadmap meeting to decide what ships against what is trending in the zeitgeist, with Anthropic and OpenAI running 10+ small releases per month that kill entire industries on launch.

Some founders are living on the edge of what's possible with AI models, shipping new features whenever model improvements enable them. Others bet their companies in public and tell stories about their ongoing transformation. A third group turns their hires into headlines. The rule underlying these three ideas is that you only earn the right to build in public if you're doing something worth talking about.

Later this month, Plane co-founder Staszek Kolarzowski (YC '17) will give a founder-to-founder webinar showcasing the AI engine he built to run sales targeting, call coaching, and commissions that scaled to $Ms in ARR. Join to see what's in production and the hard-earned lessons that shaped his systems. Register to watch live or on-demand

OpenAI will start offering early access to conversion-optimized campaigns on June 5. Accounts that have conversions set up by June 1 will have access to the feature. A screenshot of what the option will look like in the settings is available in the article.

OpenAI's Codex agentic AI platform recently received an update that introduced domain-specific workflows, a web hosting feature called Sites, and an in-place editing tool named Annotations. Sites can convert static data inputs or text documents into functional, web-hosted internal applications that users can share with their colleagues. Annotations enable more precise agent AI spreadsheet edits and updates. It eliminates full-document regeneration, which often results in broken formatting or introduced hallucinations.

Conviction's Mike Vernal argues the canonical enterprise software playbook is dead because the framework implicitly assumed calendar time that founders no longer have. Statsig wedged into experimentation before adding feature flags, session replays, and product analytics, and Rippling wedged into onboarding orchestration before stacking HR, benefits, and recruiting. Each previously needed 3-5 years to clear Act I and another 3-5 to layer the suite.

Workday spent nearly two decades getting HR and finance to speak the same language, then AI came along and changed the climate. The company spent $1.1 billion to acquire an AI interface layer to adapt to the change. Every control point in enterprise software is now operating in an environment changed by AI. Companies need to look at which points weaken in the new climate, and which ones thrive.

Some customers only stay because there aren't any alternatives.

This post looks at the patterns an investor saw inside a startup that led them to believe that investing was the right choice.

Garry Tan thinks most of the code in your AI app exists for one reason: you don't trust the model yet.

A poll tells you what people say, a market tells you what they'll actually put money behind.

A full overview of the latest batch of startups in Y Combinator's Spring 2026 cohort.