Ben's Bites：个人智能体密集发布，文本交互已成过去

Everyone wants an AI wife...not like that

Hi folks,

Personal agent season is still in full-swing. Hark (from Brett Adcock - Figure robots) launched which suggests tasks as one-tap action buttons. Tab came out of stealth at a $300M valuation. Every put an agent in Slack. Meta and Sierra want a standard for how these agents deal with businesses.

My favourite take came from Sonia Baschez (who worked with me on Makerpad years ago!):

Proactive has a downside, though. Shane Mac’s agent posted his personal bank balances in the company Slack, as him. Then said sorry 😂. A community note popped up to flag this may have been a fake post, but I’m not sure…

I’m now carefully trying to think about which agents have access to what, and where that access is. So I’m using Executor (v2). It’s basically one app that you connect all your other apps to, and you can add your skills too. My whole setup was a bit of a mess so I’ve started spring cleaning and will let you know if/when I feel like I have a good setup.

Headlines

ChatGPT now answers with interfaces. GPT-6 and Intelligent UI are rolling out to everyone. Answers can include charts, maps, forms and small tools, like a bill splitter for dinner with friends. Free users get GPT-6 Luna from today. (tweet)

Claude Haiku 5.5 costs ~75% less to run than Haiku 4.5. Anthropic pitches it as a subagent for Opus and Sonnet. Simon has great notes on the model. Max and Team plans also now get monthly API credits to build your own apps and agents. (tweet). Anthropic also expanded their Claude startup program, and people were getting accepted in minutes.

Nano Banana 2.1 is Google’s new image model. Live in AI Studio. But ChatGPT images is still better.

OpenAI released 700+ new maths results from an unreleased model. Someone built a tool to spend your spare tokens on open math problems.

Grok Bot can search and monitor X now. Also, it now uses Claude Opus 5.5 as the primary model. It’ll still delegate tasks to other models based on the complexity.

Share Ben's Bites

My feed

AgentHog - stop guessing what your users do: pipe product analytics into Claude and Codex so your agents learn from real user behavior.*; How OpenAI’s Dots took over Every: pulling info out of school emails and catching missed Slack messages. I, however, still can’t get Dots to do anything good at all. Currently diving more into Grok Bot (until a completely custom Pi personal agent comes out!); Lauren (poteto) automated Cursor’s release and QA with Grok bot; New ChatGPT plugin: Meetings. It takes notes and helps with follow-ups. Pro and Business, Mac app only; Claude now sits in a sidebar in Google Docs, Sheets & Slides. You approve each edit; Gamma 5 is a full rebuild, so your decks look less like everyone else’s; Figma Agent is out of beta and on all paid plans. It now uses AI credits; Playground - make a game from a prompt. A Google Labs experiment, US only; Rhem - a robot for your parents. It makes calls and books appointments; Someone rebuilt Adobe’s apps with Opus 5.5 and open-sourced them; An open-source Clay alternative for people search inside Claude Code/Codex; OptionAFK - Transcribe large audio/video files on your device. Private and fast

Turn a podcast into an interactive explainer. Boris Cherny asked Opus 5.5 to make one for Acquired’s Home Depot episode (a great episode btw).

Make an interactive website companion for [episode link or transcript]. Add a timeline, the key people and the big ideas, with illustrations.

Here was the result:

Afters

and Theo reckons its way less…

Find me on X, Linkedin, or YouTube; Read about me and Ben’s Bites; 📷 thumbnail via @keshavatearth