Hi there! I’m Ben, and I’m passionate about creating with agents, even though I don’t have a technical background. Here’s a collection of insights I’ve gathered along the way. If you’re eager to dive into building or enhance your ‘vibe-coding’ abilities, join our community.
Hello everyone,
I’ve been operating my own AI agent for some time now. It helps me read emails, check my calendar, manage various projects, and runs continuously on a Mac Mini at home. The entire setup was created through the terminal interface—I interacted with a coding agent rather than physically writing code.
It’s a little rough around the edges, but it’s mine and it functions perfectly. I truly enjoy it!
And I’m not alone in this endeavor. OpenClaw gained significant momentum as individuals realized the value of an always-on agent capable of accessing their information. A trend is emerging, with more people constructing personal agents, and I predict that if you haven’t started yet, you will likely jump on board this year.
This is where Dreamer steps in (and no, this isn’t an advertisement or an investment suggestion; it’s simply a tool and team I deeply admire).
David Singleton, a former CTO at Stripe and a big fan of Ben’s Bites, along with Hugo Barra, collaborated on Android. They’ve now launched Dreamer with designer Nicholas, a team of 14, and $50 million in funding.
Their pitch is straightforward: if you can envision it, you can create it.
What is Dreamer?
Dreamer is a platform that enables you to develop agentic applications through conversation. You articulate your desires, and an AI agent named “Sidekick” constructs the application for you in just minutes. For more intricate needs, there’s a detailed coding agent available. Regardless, you never need to worry about hosting or deployment; the platform manages all that for you.
This aspect is crucial for me. I often find myself overwhelmed with infrastructure tasks—getting servers up and running, ensuring everything stays functional, and troubleshooting crashes. While these tasks are valuable for learning, they divert attention from what truly matters: the projects you’re trying to build.
Over time, Sidekick learns more about you and acts as a privacy safeguard, controlling the data each Dreamer app can access. It can generate temporary agents for specialized tasks, integrate with third-party tools, and coordinate across your various applications—all set up effortlessly out of the box.
What can you actually create?
The possibilities are extensive:
-
Integrate your work and personal calendars to generate a daily briefing podcast.
-
Save and summarize links from across the web (remember Pocket? RIP) with tagging features.
-
Manage recipes that generate shopping lists and can even order items from Instacart.
-
Create a to-do list that researches and initiates tasks before you get to them.
Simply describe your needs, have Sidekick build the application, and iterate through conversation. Projects of moderate complexity take about 6-10 minutes to develop. Once completed, you can share your creations in a gallery for others to explore or remix.
Many tools are currently available for building with AI—Claude Code, Cursor, Replit, Droid. However, they still require a good deal of technical knowledge.
Dreamer is not a coding platform or an IDE with AI integrated; it’s designed to turn conversations into applications.
Current Status
After a four-month closed alpha phase filled with high engagement, Dreamer is officially launching its public beta today, in partnership with Anthropic.
While 2026 is shaping up to be the era of personal agents, we are still faced with technical obstacles. Dreamer represents the most promising tool I’ve encountered that makes this accessible to everyone.
For a deeper understanding of how Dreamer functions at its core, read David’s comprehensive exploration.
Now, let’s explore some top stories;
Peter Steinberger is joining OpenAI, and OpenClaw is being established as a foundation. He will work towards integrating agents that interact seamlessly with one another into OpenAI’s main products.
Recently, a series of new AI models were launched:
OpenAI’s GPT-5.3-Codex Spark – offers speeds 3x-5x faster than GPT-5.3-Codex. Think of it as a streamlined model (though it does experience some performance trade-offs). This text-only model features a 128k context window and operates on Cerebras’ hardware, accessible via Pro subscriptions ($200/month). See it in action with Pi.
Minimax M2.5 and GLM-5 – two noteworthy models emerging from Chinese laboratories. M2.5 performs comparably to Opus 4.5 in coding benchmarks, while GLM exhibits impressive tool-calling capabilities—all at a significantly lower cost than Opus or GPT models.
Gemini Deep Think 3 – built on Gemini 3 Pro, achieved an impressive 84.6% score on ARC-AGI 2 (compared to Opus 4.6’s 68.8%). Available for Gemini Ultra subscribers, details on its API availability are limited, yet it excels in academic testing compared to other models—as noted by Google/DeepMind.
Interestingly, OpenAI claims that GPT-5.2 made strides in theoretical physics. It simplified a complex formula concerning a particle’s behavior, which, while not groundbreaking, denotes progress of “new work.” OpenAI is also exploring other complex mathematical problems, such as those presented on 1stproof.org (find out more).
Why do you always find a meeting bot in your Zoom calls? The credit goes to Recall.ai. They provide the backbone for every meeting AI application, including Cluely, Hubspot, and Clickup. Start today with $100 in credits.
From the portfolio:
-
The vibe coding trap – Max argues that Software as a Service (SaaS) isn’t dead. Easy development doesn’t always equate to easy ownership and maintenance.
-
Andreas created a video with a similar perspective (it’s currently one of my most trusted sources of insight) – as one-shot software becomes the norm, what should founders focus on? Paul Graham also addresses this concept in his essay, “Taste for Makers” from 2002.
-
The ten moats of vertical software (and their relationship with large language models).
-
Demoing the non-existent AI computer – envisioning software that lacks applications, incorporates generative features, enables local inference, and emphasizes personalized experience.
-
How system prompts influence agent behavior.
-
A running diary of achieving real outcomes with OpenClaw.
-
Why I don’t worry about job loss due to AI: ordinary people will likely remain unaffected.
-
Changes in structure within an AI-dominated software landscape.
-
Dario Amodei’s appearance on Dwarkesh’s podcast: a dialogue about revenue generation as a frontier lab and the feasibility of simply acquiring more GPUs if AGI is imminent.
-
Is it feasible for AI agents to produce polished boardroom-ready reports based on real, fact-checked data? Yes, it is with Superagent. Pose a question, and it dives into action: Subagents research the topic, explore credible sources, and compile everything into reports, slides, documents, or websites.*
Share
That’s a wrap for today! Feel free to drop a comment and share your thoughts. 👋
* Special thanks to our sponsors who made this newsletter possible 🙂
Interested in partnering with us in Q1?
