Fantastic write-up on the intentional design decisions for Grok Bot. It's a great view into some of the hard tradeoffs to be made as we transition to the agent era of software. Highly recommend reading it! My own thoughts: 1) The first thing I appreciated is the attention to the details, you can tell the team sweated the small things in the right ways. Building things, feeling them, refining and iterating until it felt right to them. 2) This design process is a testament to product-building as a truth-seeking journey. You can't be afraid to be wrong when opinionated. But you also can't expect to always be right with every little bet you're making. There are unavoidable hard tradeoffs, but building product is committing to the journey of orienting around a "truth" that you feel is right. I've always believed the best products reflect this point of view, guided by an insight in what users need. 3) Tangibly, this shows up in many places with Grok Bot. I liked their focus on reducing the number of primitives that users have to confront (the 5 are bots, chats, prompts, tools, artifacts). I'm a big fan of this push bc I believe crossing the chasm from chat to the agent era for the masses, requires an approachable bridge -- familiar enough to take the first steps, robust enough to support you across to the other side. Personally, I think routines/skills could still be simplified in product execution, but overall, I like the direction of wrapping them under prompts as the global primitive. 4) The central primitive is a Bot, and there's already been a philosophical debate brewing about single agent vs. multi-agent as the primary pattern. Do you want to talk to Jarvis and let him handle everything underneath, or do you want to create a team, an army of intelligent digital entities that work together and relentlessly execute whatever you hand them? I think this is an open debate, and a good one for anyone building agents to have strong POVs on. Grok Bot intentionally took one side, where they believe the future of work can continuously be organized under multiple bots just handling things. Town AI took the other approach, with 1 assistant that can handle everything work-related and very clearly anthropomorphized. Instinct is similar although still opinionated and feels familiar, without explicitly coming across as a human. All these products are starting to instantiate agentic presence, and form new behaviors on how humans collaborate, delegate, and trust agents with more. There's one argument that reinforces Town AI, in that humans are most familiar with working with other humans and so the near-term product should index on familiarity. There's another like Grok Bot which in name + design practice, seeks to make it dead-simple to talk to Bots, even though they are a net new entity. And then there's the case of Instinct, which toes the line by doing a bit of both: a single agent that feels familiar because it's already in iMessage, but is still a new AI-human interaction pattern. You can have a point of view on each, why it's good/bad, and varies depending on the person and what they are trying to do. Maybe the same person prefers a single anthropomorphic agent for personal life, and a team of digital employee bots for all work. Or vice versa. Maybe it depends on whether you work in tech or are tech-savvy. Maybe it depends on whether you've managed people before or not. The point is, it's good to be strongly opinionated about these directions because that is literally going to drive all the decision-making about the details of your product. It will make or break the experience. And, it is all still early and no one can predict how we navigate forward so you might as well take your shot and go hard at it. 5) The last thing I'd call out is the closing paragraph on the central question: "did this help someone delegate, or did it give them one more thing to manage?" I strongly believe we are haded towards disappearing interfaces, where we unlock more power by taking more things away from human eyeballs. This should be our north star: we're creating more software in aggregate, but we're delegating more than we ever have. The invisible layer for agents, the visible layer for chosen attention.

If you don’t like personal agents, this could be why Although I think automation is only one of the benefits of personal agents, and the first one to show real value. There will be far more delightful benefits too, and maybe then you will see it

unfortunate and will not age well. there are many risks and concerns and I would not sweep any of them under the rug but in no version of risk/reward is a ban the right tradeoff to make this is unequivocally a mistake

This is an important question on willingness to delegate important tasks to personal agents I have two thoughts: 1) I believe early adoption in these use cases requires a certain belief, a certain leap of faith to trust an agent to do something slightly beyond your comfort zone of what you think it can do. I know to many, this is not an acceptable answer given the high bar for trust and the nature of certain tasks that are quite expensive to fix. There is some judgment one must exercise in what they’re testing these products against, but overall — everything is early, and I believe you should calibrate your expectations against the reality that things might not work perfectly but it’s in the effort of “pushing” that we go from something rough to something great. Most important point here is that it’s a collective effort: it isn’t just on the founders or builders to nail these products, building a great agent for everyday life will require massive contact with reality and it’s also on all of us (the real world) to play our role and push these builders and products to actually meet the hype. That’s how I view my role. 2) this question has a certain framing on how you measure your time and time saved that I think is quite loaded. 5 mins checking the work of an agent != 5 mins filling out the form itself. Even if there was an absolute time savings in how much faster task execution is in the hands of an agent (which there often is), which remains even after you factor in the verification tax, I would still argue delegation is worth it for me because it’s time saved in the context of my real life. Those 5 minutes of checking the agent’s work later actually save me 5 minutes earlier in the day and allow me to be present for whatever else I want to do, so I can’t measure these two things apples to apples. There’s an opportunity cost to allocating time in moments of the day, which I think is one of the true unlocks in task delegation to agents. I am biased as a new parent where I can measure time in these ways, but I expect more people will intuit the same tradeoffs as we hand off more to AI.

What a joy it is to set goals and have dreams, to aspire and fall short, to struggle and withstand, to persevere and overcome, to travel a long distance a single step at a time To look back on the road traveled, the road ahead, and to say: “that was something” And to set goals and have dreams, and do it all over again

Instinct won’t end up like Clubhouse because at the very least, it will get bought by a frontier lab The real world failure modes it’s already getting from users is a richer eval set for RL post training than anything the labs have A beautiful long-tail of tasks that teach the model on how to be a better assistant, because it’s made contact with reality

1. It’s great to see the spotlight on @nikolasklein — he’s a fun, caring, uniquely n of 1 builder. IMO: he represents the future of building, flexing across high craft design, product taste, and skillful storytelling 2. Listen carefully to the point. It’s not just about the canvas as a substrate, it’s about a way of working and thinking that is undervalued in the current moment. If LLMs gain intelligence through compression, there is a different type of intelligence to gain from the opposite motion. We need divergence and convergence at different points to arrive at something that lasts. It’s the way the creative process has always been. Building software that endures will look more like a creative act than an assembly line. Not obvious enough yet.

People in tech are so obsessed with booms and busts and debating what’s hype vs real But history has always told us how to decide what to pay attention to: breakthrough moments, where something that seemed impossible became somehow possible. The self-attention mechanism in transformers is gonna go down as a defining moment for the rest of this century. And we are still so early in extracting real world value from it. There is so much to rebuild, our entire relationship with software needs to evolve to assuming we always had Gen AI and that the cost of digital creation will effectively be zero. Websites, apps, videos, images, sounds…all of it and the entire largely static ecosystem built around exchanging this information online through software needs to be reimagined. Some will experiment incrementally, some will more aggressively reimagine. But either way, we won’t have more of the same