Zuckerberg’s Wonder: Ben’s Bites

Welcome everyone,

As a quick note: I’m developing a resource for non-developers who may be unsure about the potential of personal agents. Do you converse with one yet find yourself puzzled by the enthusiasm surrounding them? If this resonates with you, please respond to this email. I’m seeking about 20 individuals to help test a new concept, and I’ll add you to a dedicated Slack channel.

Now, let’s dive in. Yet again, there’s a debate on whether software is dead. These discussions seem as common as opinions—everyone has one, myself included.

Rest assured, software is alive and kicking. We are likely witnessing unprecedented levels of code being written. Something might be obsolete, but we’re still figuring out what that is.

Personally, I utilize far fewer SaaS tools now than in the past. My personal agent, Grok Bot, for example, reviewed our business subscriptions and helped me save $6,000 annually by eliminating services like Webflow, Resend, Cal, and Zoom, among others. When I consider new tools, I primarily ask if my bots can work with them. I no longer need to navigate interfaces—now it’s simply software communicating with software, also known as ‘headless’ operations.

Moreover, I’m constantly launching websites for a range of purposes. I recently created a site to evaluate all my subscriptions, making it easier to decide which ones to keep or cancel. Once I accomplish my objective, I dispose of the site—though realistically, it usually languishes untouched for ages.

The primary challenge I’ve noticed is that many individuals don’t perceive themselves as ‘builders.’ They often fail to consider creating sites for various tasks—like managing to-dos, email sorting, meal planning, or family schedules. But for me, constructing these resources is second nature. Ironically, I prefer visual content over blocks of text any day.

I believe that personal agents will empower everyday users to create software, as long as it feels effortless. Currently, many of these agents are somewhat proactive, presenting users with tasks to complete (aren’t they supposed to be working for us?!) instead of solely delivering software upon request.

If you were given a personalized site to manage your to-dos, I bet you’d want to adjust the colors, layout, or something else. Back in the MySpace days, we customized our profiles with music, themes, and layouts. Some even learned HTML & CSS to take customization further.

Software creation should feel like enhancing something already yours.

Here’s a list of additional concepts I’ve been pondering:

• Sloppy coding is unfairly criticized. Developers often criticize us for producing messy code with our agents. But if it functions well, is secure, and performs adequately, who really cares? Certainly, it matters more for services that provide critical infrastructure to paying clients. However, for personal tools and sites, messy code suffices. I believe it’s the bot’s responsibility to manage those inconsistencies. If my code isn’t perfect, that’s fine—fix it!

• Messiness leads to improvement. Remember how developers got their start? By writing code that was less than perfect. The objective isn’t to master coding from the get-go. If you want to build something, bumps in the road are inevitable. Both developers and non-developers improve by practicing and learning more about the system over time.

Below are my sources that inspired this discussion:

Muse Wonder and Wonder Full – fresh models from Meta. Wonder is positioned as a cost-effective alternative in the Fable/Astra range ($4/$20 per million tokens compared to $10/$50 for Fable and Astra). It can be accessed through Meta’s API, available in Muse Code, and soon will drive Muse (Meta’s personal agent) as well.

Wonder Full delivers even better performance and falls just short of Opus 5.5 standards. It’s currently exclusive to the Muse Code subscriber plan (starting at $50/month).

In my pursuit of these models, I’ve subscribed yet again to an AI service. My early tests with Wonder (medium) and Wonder Full (high) revealed that the medium option was more verbose than the high, though both seemed sluggish with poor UI. I didn’t apply them to ‘proper’ coding tasks and might not return to these models unless I’m exclusively testing them again.

Zuck and Alexander Wang appeared on the Cheeky Pint podcast discussing their journeys into new realms.

Claude Code Projects are now accessible to all Pro and Max users. The format organizes everything into one master chat to submit requirements, ideas, and tasks, with Claude generating new threads to tackle them.

A Unified Plan for Cursor, Grok, and X – Four subscription tiers ranging from $8/month to $200/month. Increased usage for Grok and Composer models, with significant limitations for Claude and GPT models. Additionally, Grok Bot now has its own email. I even got mine to register for its own Gmail account 😊

If your business is on the decline, there may still be a $1M+ opportunity. Explore how micro1 compensates qualifying firms for licensing operational data that assists in training AI models about workflows. Discover if your business is eligible here*

Share Ben’s Bites

Instinct’s founder Noah Shinn shared this prompt, which can be utilized by any agent that has email reading capabilities.

Sarah’s essay on the end of friction is a must-read.

sarah guo@saranormous

the impact of personal agents, the end of friction, and the economic effects of finally taking action can be explored further here: https://t.co/jEOtEaFvbx

2:55 PM · Oct 12, 2026 · 128K Views

77 Replies · 46 Reposts · 611 Likes

My current favorite personal agent is Grok Bot. Unfortunately, I haven’t been able to test Muse in the UK (despite trying various methods). Grok Bot operates quickly, is quite capable, and has a friendly interface. Dots requires some refinement before I give it another try, but I’ve already reached out to OpenAI regarding my feedback 😬. I’m attempting to utilize Grok Bot for as many tasks as possible to assess its performance and identify any weaknesses. I create new bots for each specific task and archive them upon completion.

Recently, I reviewed 195 domains across three different platforms and managed to reduce my auto-renewal expenses from an astounding $4,000 a year to half that. I’m also consolidating them onto one platform.

One of my bots has taken over my email triage and is currently handling labeling and archiving tasks, although it needs some fine-tuning.

As is often the case with these tools, the best strategy for now is to dive in and utilize them for as many applications as you can imagine.

Yet, Grok Bot has its quirks—sometimes it can only access my computer if I manually remind it and ensure I’m on my computer rather than my phone (which seems to be a known issue).

Sources that fueled this discussion:

• Software is dead

• Vibe coding and slop code

Nader Khalil🍊@NaderLikeLadder

Introducing San Theft Auto: a racing game set in San Francisco—multiplayer and open-source! I spent 950 million tokens having GPT-6 meticulously map the city in Blender. You can try it out here: santheftauto.com

6:12 PM · Oct 10, 2026 · 36K Views

56 Replies · 23 Reposts · 349 Likes

lauren@poteto

What did you accomplish with @bot this week?