this is a new format i'd like to try

this is an example of a new format i like the idea of… the original of this email (in our current format) is here. please vote at the bottom of this post which you prefer. [this top section is where i’ll talk about things on my mind. sometimes it’ll be tool comparisons, sometimes explainers] A founder I’m working with asked me yesterday if I’d read Sequoia’s post; Services, the new software. I had when it first came out, but I re-read it. TLDR... AI companies that sell outcomes (autopilots) rather than tools (copilots) will capture the far larger services budget. Example: Harvey started selling copilot as a tool to law firms but is moving towards completed contracts and filings as an outcome. The best strategy is to start with outsourced, intelligence-heavy tasks as a wedge before expanding into higher-judgement work. It made me think of the vertical agents that we are seeing today, which effectively start as custom skills for one task. It could be documentation writing, animated videos, explainer visualisations, email triaging, sales reporting, etc. We are seeing it over and over again now with agents for CFOs, CEOs, CMOs and so on. We will see it for every profession and job level there is. This was on my mind as I saw the Manicule launch - “an AI native technical documentation studio”. They have packaged up an agent+human services into something that feels like consulting/contract work. I think it’s smart (not just because it links with the Sequoia post I promise). I think the packaging is what’s smart (tbd on the job-to-be-done they are solving - won’t agents just write great agent docs for agents to read?). Another founder I spoke to yesterday (I think I’m going to invest) is working on specification and planning for development workflows. We’ve all heard ‘you should use plan mode before coding’ and the more I see people build/fork their own orchestration platforms the more planning feels like the only step to truly matter. If you could have your team (+ agents) all collaborating on the specification and spending time thinking about tradeoffs, implementation details and the like, the agent should be able to execute that. All of the grunt work is done in planning. And when code is up for review, it’s much easier to read and make sense of the spec that created the code than trawling through the code itself. And broadly, code review feels like it could be a thing of the past when models hit a certain threshold. tbd. Codex now has 2M+ weekly active users, and OpenAI API use is 20% up since GPT-5.4 was released. Subagents are also now live in both the Codex app and CLI Manus (recently acquired by Meta) launched a desktop app, My Computer, to compete with Codex/Claude Code/OpenClaw, etc. Jensen Huang says Nvidia expects to generate $1T+ in sales from its flagship AI chips through the end of 2027. They also released NemoClaw Travis Kalanick of Uber fame is back with a bang. TBPN video interview The socio-cultural importance of the Claw movement: Lobster Boil Nebula. The best packaged OpenClaw. Customise your own ‘Claude for Chrome’ extension - AMA with Mario here (~38min mark) Cal (the better version of Calendly) launched its own agent and skills. HubSpot launched HubCode - create apps on top of your HubSpot data. Andrew Ng’s new project context-hub lets you fetch curated documentation and annotate it for future use. A terminal-first twitter CLI where you can get for-you and following feeds, bookmarks, etc. more general-purpose version of Karpathy’s ‘autoresearch’ Context7 just released their CLI for code docs a memory plugin for OpenClaw Read about me and Ben’s Bites 📷 thumbnail by @keshavatearth * sponsors who make this newsletter possible :) Wanna partner with us for the next quarter? Email us at shanice@bensbites.com or k@bensbites.com