Last updated May 27, 2026
If this month has one word, it is busy.
DreamNum has taken up most of my time. The working hours are long, and the shape of my day has changed around that. There is less loose time now. Less time for side projects, less time for cycling, less time to let ideas wander without turning immediately into work.
That is not all bad. The work itself sits close to something I care about: making agents better at using real software. But the cost is also obvious. A month can disappear quickly when one project expands to fill almost every usable hour.
At DreamNum, my main focus has been a simple but difficult question: how do we make agents understand and operate spreadsheets better?
Spreadsheets are not just grids of values. They carry structure, intent, dependencies, formatting, implicit workflows, and a lot of human context. Most traditional programming interfaces treat them as files to manipulate. That is useful, but it is not necessarily the most natural surface for an agent.
Recently I started designing a new DSL for describing spreadsheets. The goal is to give agents a representation that is more native to how they reason about tasks: what the sheet is, what parts matter, how ranges relate to each other, what actions are safe, and what changes mean. It feels more agent-friendly than asking a model to drive a Python library directly.
This is still early, but it feels like the right layer. If Univer is going to be spreadsheet infrastructure for agents, then the interface cannot just be a human spreadsheet editor with an LLM bolted on. Agents need their own language for the workbook.
Alan has also crossed another line for me.
I built Alan for macOS, and it has become the terminal I use every day. It has replaced Warp in my own workflow, which is not a small thing. A terminal is not an occasional tool. It is where I live for a large part of the day. If Alan can take that spot, it means the product is no longer just an agent runtime idea. It is becoming part of my actual computing environment.
The agent side has not moved forward as visibly this month. Most of the progress has been in the macOS terminal surface and daily usability. But the next set of ideas is starting to form. One direction I keep thinking about is inspired by Thinking, Fast and Slow: giving Alan a System 1 / System 2 model, where it can normally act quickly but deliberately escalate to deeper reasoning when a task deserves it.
That is still more design than implementation. But it feels connected to the deeper thing I want Alan to explore. I do not just want an agent that can call tools correctly. I want it to show something closer to a human-like behavior pattern: quick intuition, slower reflection, memory, habits, and a sense of continuity.
“Silicon-based life” is a loaded phrase, but it points closer to the feeling than “chatbot with tools.” I want Alan to behave less like software waiting for commands, and more like a being that can act with its own recognizable rhythm.
openrouter-rs is in a steadier phase now.
The work is less dramatic than the 0.6.x and 0.8.x milestones. Most of it is weekly maintenance: checking the automatically generated drift against the official OpenRouter API, then adjusting the SDK when the official surface changes.
That is a good place for the project to be. It does not need to prove itself every month anymore. It needs to stay correct.
My cat has been eating a lot more recently, which should be comforting. But he is still losing weight.
The doctor said this is a normal pattern with cancer. I understand that clinically, but it does not make it easier to watch. He still vomits sometimes. I still give him medicine every day. I still do not know how much it is helping.
There is no clean engineering problem here. No satisfying debugging loop. Just a living thing I care about, getting thinner even when he eats.
I am anxious about it.
Another thing changed in May: I stopped biking to work.
Shenzhen has entered summer, and it is simply too hot. The commute that helped clear my head in April became hard to sustain. Since stopping, I can feel my body getting worse again. That worries me more than I expected.
Work is taking more time. Cycling disappeared. My body is giving me feedback. None of this is surprising, but it is still uncomfortable to see the pattern so clearly.
There is also pressure at work. My boss thinks I may not be focused enough on the company’s project and wants to see more output there. That bothers me. Part of me understands the expectation. Part of me also feels stretched thin already.
I do not have a neat answer for that yet.
June should mostly be about the spreadsheet DSL.
I want to push it from idea into something usable: not just a clever format, but a real interface that helps agents read, reason about, and modify workbooks. If that works, it could become the center of what I am building at DreamNum.
Alan will continue in the background. openrouter-rs will keep being maintained. Sealbox is still paused, though I now suspect it may eventually come back as a secret management and usage tool for agents. Kira is fully gone. Alan has taken its place.
For now, the goal is simpler.
Stay alive, and keep going.