Key Takeaways
- AI agent developer tools are shipping rapidly: GBrain personalizes agents for Codex/Claude Code, Perplexity Computer adds email tasks and permission controls, Replit launches black-box pentesting and enterprise governance.
- Vercel bets $1M on Sandbox security verification, inviting public attempts to escape any model.
- Agent infrastructure expands: Hermes Agent reintroduces Bot Mode, AgentJourney lets teams test agent readiness, and River API claims RL performance wins over Tinker.
- Real-world startup signals: TryNearby reports 120+ paid restaurants with >130% growth, and 776 hits a $10M milestone.
- Policy and debate: Ro Khanna faces criticism over financial disclosure and stock trading; the "GPT wrapper is useless" narrative is debunked as dozens of such companies reach billion-dollar valuations.
1. AI Agents and Developer Tools
Garry Tan's GBrain introduces a personalized agent generator for Codex and Claude Code: a 12-question setup creates a SOUL.md and installs 70 skills, integrating with any AI harness and Postgres+pgvector to give users ownership of their own AI agent memory and skills. — via 1 2
Perplexity's Computer adds two major capabilities: connector permission controls (Allow / Always Ask / Deny) with single-approval or thread-wide availability, and email-based task execution by sending mail to [email protected], with the same audit trails as in-app tasks. — via 1 2
Replit now supports black-box penetration testing to find vulnerabilities that code scanning misses, with Replit Agent able to fix issues in one click; it also announced enterprise governance tools including full audit logs, workspace controls, and admin APIs. — via 1 2 3
Vercel is committing $1M to publicly verify the security of Vercel Sandbox, inviting anyone to attempt escape against any model; findings will be fixed, iterated, and shared publicly to strengthen global cyber safety. — via 1
Hermes Agent reintroduces "Bot Mode" as a session alternative: each bot profile has its own chat, tasks, description, avatar, and memory, can communicate with other bots, and maintains its own skills, tools, and external connections. — via 1
Agent tooling grew: AgentJourney lets anyone run an intent on any site for free, showing the agent's journey, token cost, and pricing in real time; Monid Discover enables agents to search 1,300+ APIs and tools, compare vendors, and call/pay at runtime. — via 1 2
River API outperformed Tinker in reinforcement learning tests under identical training code, with the team emphasizing routing replay and other details to deliver best results. — via 1 2
2. Business and Product Milestones
Basecamp 5 shipped a refreshed My Bar with My Activity, letting users quickly see recent activity, return to recent actions, re-enter conversations, or review work; it also features a subtle screen-bottom reminder in the 15 minutes before an event, turning greener as the start approaches. — via 1 2
Alexis Ohanian is building a "women's sports Reddit" product and thanked alpha testers, listing new features: drag someone into starter/bench/waive, full court stats, @mention alerts, and auto-generated post-game roundup posts. — via 1 2
TryNearby (per Amjad Masad) has 120+ paying restaurants in Southern California, growing over 130% since batch launch with >90% retention—and achieved AI-level growth without even mentioning AI in its pitch. — via 1
776 (Alexis Ohanian's fund) hit a $10M milestone; Ohanian reminded founders that "overnight success" is never overnight and you have to build a product people love. — via 1 2
Garry Tan shared that his company was rejected by Y Combinator four times before focusing on product and users; a self-run Launch Week went viral, and YC eventually reached out to admit them. — via 1
3. Policy, Society, and Industry Debates
Ro Khanna faces double criticism: Naval highlighted that the congressman again filed a 353-page paper financial disclosure (unindexed and unsearchable), and traded on 244 of 250 trading days last year (~22 trades/day) despite claiming to lead a congressional stock-trading ban; Paul Graham added his response reads like AI-generated content. — via 1 2 3
EU packaging rules may inadvertently kill the single market: businesses shipping packaged goods to other EU countries must register in each of 27 states, costing up to £1,000 per registration with no unified mechanism—described as "detonating a nuclear bomb in your own house to stop a thief stealing spoons." — via 1
The "GPT wrapper is useless" narrative from a year ago is now the biggest lie, says Greg Isenberg, with dozens of such companies now worth billions and hundreds worth $10M+; he hints there's a lesson in that reversal. — via 1
Agentic commerce hasn't hit its breakthrough yet, while a "build everything" mindset and taste at scale dominate current agentic coding discussions, per Patrick Collison. — via 1
Critique of wealth taxes (via Naval) argues unrealized gains taxes rely on unreliable private-company valuations and ignore illiquidity, weakening founder incentives; Amazon and Nvidia show successful firms create more jobs/taxes than founders' personal wealth. — via 1
