2026-05-01. Nomoyu Daily for Indie Developers (Issue 349)
📰 News
Beyond model rankings, the battlefield for agents is also here
If you are still asking “which large model is the smartest,” you may already be half a step behind.
In this interview, Notion reveals a colder reality: the outcome for AI agents is not decided inside the chat box, nor on parameter leaderboards. It is decided by who controls the actual workplace.
The model is only the engine. What really determines whether the car can get on the road is the road, traffic lights, brakes, insurance, and driving rules.
Agents are not a new species of human. They are new workflows.
On the surface, Notion’s Custom Agents let AI help you read email, create databases, research information, and assign tasks. It does not sound flashy. It may even feel like “another wrapper.”
But do not be too quick. In the AI era, the word “wrapper” may be seriously underestimated.
AWS has CloudWatch, yet Datadog still lives very well. Enterprises do not want merely “a monitoring feature in the cloud.” They want someone who truly understands how engineering teams observe systems, troubleshoot problems, and collaborate.
By the same logic, model companies have capabilities, while Notion has the work scenarios. A scenario is not a skin. A scenario is power.
The sharpest line in the interview is this: agents do not replace people; they replace workflows.
A bug appears in Slack. An agent decides which team owns it, creates a task, and replies with a notification. A meeting ends. An agent organizes the notes, generates to-dos, and sends follow-up messages. An application email arrives. An agent reads the content, checks the background, and updates the database.
This is not “AI helps me write a paragraph.”
This is invisible office administration, PM assistance, and operations coordination being swallowed by software.
Real agents must first be put in a cage
But mature agents are precisely not better when they are more free.
Notion team’s view of MCP and CLI is interesting: CLI is powerful because agents can create and fix tools in the terminal; MCP is plain because its permissions are clear and it can only call given tools.
Behind this is a counterintuitive point: enterprise AI is not pursuing “it can do anything.” It is pursuing “it can do what it should, and cannot do what it should not.”
An agent that can open every door is not productivity. It is an accident. Only an agent disciplined by permissions, cost, and tool boundaries can run in enterprise backends for a long time.
Even sharper is the view on cost.
If a deterministic API can solve a problem, and you still make an LLM repeatedly read context and call tools, that is not intelligence. It is waste.
Being able to use models is not rare. Being able to make models appear less often is the real product capability.
Software engineers will not disappear, but they will be rewritten
The core capabilities of the next generation of AI products are not mystical system prompts, but three things: tools, evaluation, and permissions.
Notion tried XML early on, then found that models did not like it, so they moved to Markdown. Database queries are also made as close as possible to the SQLite style that models already understand.
Do not make models adapt to your internal world. Translate the world into a form models can understand.
This is also why they do not worship the idea of “training our own large model.” When tools change every day, retraining a model can actually slow things down.
The smarter approach is to build the outer system well: tools can be swapped, failures can be evaluated, costs can be calculated, and permissions can be tightened.
They are even building “Notion’s final exam”: not to prove how smart a model is, but to find the boundaries where the current agent is most likely to fall.
Future software engineers will not simply disappear. They will be forced to move up a level.
Before, you wrote code. Now you need to command a group of agents that can write code, run tests, open PRs, and also make foolish mistakes.
What is scarce is no longer “people who can type.” It is people who can design closed loops.
Whoever records the work rewrites the work
Meeting notes follow the same logic.
They are not just speech-to-text. They are creating a new enterprise data layer. Every discussion, disagreement, decision, and to-do becomes context that future agents can search, cite, and execute.
Do not underestimate this change.
Whoever controls enterprise memory controls the material agents use to judge. Whoever controls the collaboration entry point controls the last mile of AI execution.
So stop understanding agents as robots that are better at chatting.
They are more like breathing workflow systems: with memory, permissions, tools, evaluation, cost awareness, and self-repair after failure.
Models are important, of course, but models are becoming electricity.
What is truly valuable is who can wire that electricity into the factory, make the machines run, and avoid burning the factory down.
Here is a criterion for every young technologist: when looking at AI products in the future, do not only look at which model they call.
Look at whether they have a real worksite, a data flywheel, a permission model, an evaluation system, and whether they actually free people from workflows.
The future will not be one omnipotent large model swallowing everything.
The future is more likely to be countless small agents quietly moving things between your documents, emails, meetings, codebases, and databases.
They will not show off, but they will gradually eat away at the most annoying, fragmented, and low-value work of today.
If the entrance of the PC era was the desktop, and the entrance of the mobile era was the app, then the entrance of the agent era is the system of record.
Whoever records the work can rewrite the work.
🖥️ Software
XploitScan
XploitScan is a security scanning tool for AI-generated code, capable of detecting vulnerabilities such as hardcoded secrets and missing signature verification.
Vantage
Vantage is an aesthetically refined all-asset management tool that supports US, Hong Kong, and A-share stocks, funds, crypto, and other asset types, with local-first storage and intelligent OCR recognition.
Folderlister
Folderlister is an eBay listing assistant based on voice and image analysis. It supports describing item information by voice and automatically matching images with data.
Adrio
Adrio is an AI-based e-commerce ad creative generation tool that supports ad cloning and personalized optimization, improving ad production efficiency.
Deskboard
Deskboard is a free local app that turns folders into customizable visual boards, supporting file opening, notes, and to-do items.
DelayCam
DelayCam is a mobile app for iOS and Android that enables instant replay of movements through a camera with adjustable delay, suitable for training scenarios such as golf, dance, and weightlifting.
Aftercut
Aftercut is a screen recording tool positioned as an alternative to ScreenStudio. It supports paid exports, has reached $800 in sales, and plans to launch subscription-based premium features.
Cronolog
Cronolog is a private timeline app for recording and visualizing personal life events. It supports fuzzy date input and helps users clarify the course of their lives.
SummAgent
SummAgent is a Chrome extension designed to help users save time, focusing on productivity improvement and AI assistance.
🎮 Games
Unsatisfactory
Unsatisfactory is a comedic digital mind simulation game that supports emotional processing and system upgrades, offering abstract environment exploration in 15 to 20 minutes.
NeonPaths
NeonPaths is a relaxing mobile puzzle game with daily puzzles and multiple modes, featuring progress tracking and sharing.
Gems
Gems is a browser-based strategic matching game that combines Bejeweled with Slay the Spire. It supports mobile devices and offers unique gameplay without ads.
🌐 Websites
BrewCard
BrewCard is a free coffee recipe sharing website that supports custom designs and public link sharing.
✍️ Notes
Daily project information:
Website: https://www.nomoyu.com/
RSS: https://www.nomoyu.com/rss/rss.xml
WeChat Official Account: 明航的AI副业
Feel free to connect and exchange ideas
See the website for all links.