It is now early morning on April 15, 2026.
Recently, I've advanced the development of auth-mini to version 0.2.0. It's basically usable now and has been officially deployed at https://auth.ntnl.io. From now on, it can handle the authentication functions for our various new systems.
I developed this entire project using Agentic Engineering.
Additionally, the SOAR project is steadily progressing. This is the first time the SOAR project has been revealed; it's the project name for the Capital Persistence War, deployed at https://soar.ntnl.io. Everyone is welcome to take a look.
This SOAR project was also developed using Agentic Engineering.
All of these were developed by GPT-5.4. It has probably consumed tens of billions of tokens in total. On average, I can burn through 100-200 million tokens per day, with the highest 24-hour period reaching 480 million.
For me, the more surprising discovery is that I only need minimal prompts for each task to allow the AI to naturally generate many details, and their quality is already sufficiently high.
It can achieve about 80% quality. I think that's enough.
Under my prompts, it can now independently complete a full work cycle: creating a worktree, baseline testing, exploration, development, testing, submitting a PR, addressing review issues, merging the PR, and cleaning up the worktree.
However, it's crucial to note that this process still relies on TDD (Test-Driven Development). We need to set up CI/CD pipelines on the GitHub repository.
I believe this Agentic Engineering development model has reached a very good state. Its efficiency far surpasses traditional development.
But, I think it's at most equivalent to a Senior SDE individual contributor. It no longer needs me as a Mentor constantly correcting it on the side, but it hasn't reached the level of a Dev Team Leader yet. It still can't manage, plan, and break down tasks very well.
The real question is, how can we get AI to take on the work of a Team Leader?
We can start by discussing how to improve concurrency. In current AI development, an Agent Session typically contains 4-12 Tasks and runs for 1-2 hours, or even longer. Over long periods, it can even encounter automatic context compression issues, leading to quality degradation. For a single task, it modifies the interface first, then develops the backend, then the frontend—all sequentially.
If we can improve concurrency, the AI could work on multiple tasks simultaneously, greatly boosting efficiency.
Unfortunately, the Session layer in OpenCode doesn't support this operation. We can't achieve this through Skills, Plugins, etc. We must develop a scheduling system above the Session layer to implement this functionality.
We need to treat an OpenCode Session as a basic unit for task execution. Each Session can be bound to a Task for execution. We need to develop a scheduling system above the OpenCode Session layer to manage the lifecycle of these Sessions.
The work cycle of a Worktree + PR mentioned earlier is a Task Session.
But besides that, when starting a task, we still need to clarify the problem through an interview and then break it down into several Tasks. This interview and breakdown process is also an OpenCode Session, but it doesn't require opening a Worktree or a PR, so it's not a Task Session.
I can start a new project called AIM, Agentic Intelligence Manager, specifically for this scheduling system. This is an AI-Native project. It can enforce the binding of OpenCode + Git WorkTree + GitHub PR to form that work cycle.
Furthermore, I want AIM's state management to be handled by a SQLite database. This database doesn't require permissions; it's essentially just a file. Its sole purpose is to coordinate between Sessions, recording Session status, task status, dependencies, etc. It doesn't need to provide any external interfaces.
The AI can directly read and write to this database through scripts. I don't need to design any APIs; that's unnecessary. I just need to define the SQLite schema. The AI can directly operate on the database using SQL.
Even if a Task Session gets interrupted, I can arrange for an AI to automatically push a prompt to that Task Session to let it continue execution.
This thing shouldn't be very large. Once it's done, my daily token consumption might increase to tens of billions, and my daily output could rise to dozens of PRs.
Let's give it a try. This leverage is important.
