trine: notes by zerone0x
Notes on AI-agent runtimes, developer tools, and what happens at the edge of shipping software. Who writes them, and what I build, is on trine.dev.
Featured
On Units: The PR and Its Author
About 180 PRs across four projects, written and verified by agents. The first week is easy. After that the unit isn't the commit or the week, it's the PR, and the author is a group of PRs that needs a credit score.
On Self-Improvement
Notes on letting an agent improve its own harness. It's almost always overfitting, the patch is the cheap part, and everything that works has a benchmark under it. We have an inbox.
On Routing
Routers and gateways are one pile of dirty work, seen from two ends. The hard question isn't which model — it's what you're routing. The open one is what 'compatible' even means.
On Harness
Six articles saying the same thing: move yourself out of the loop and onto the gate. And why three of them were never graded by anyone.
Self-Evolution Is a Data Problem
Everyone wants agents that improve themselves. The blocker isn't intelligence — it's data. Observability first, a centralized fleet second, the agent last.
FDE: The Job Is the Gaps
Confirm the requirement, build on the product, deliver. Accurate and useless — the real work is extracting specs nobody can state, winning access, and shipping evidence instead of demos.
Agent Harness Turns Development into Executable Patterns
Agent harnesses are not just multi-agent wrappers. They route context, package workflow skills, expose project-specific tools, and use traces as the feedback surface for improvement.
The Memory Stack
Agent memory isn't a switch you flip. It's a four-layer stack: a provider, a policy MCP, the laws you write, the router you fork.
What I'd Steal From Hermes for the Next Orbit
Orbit was my first stateful agent harness — scout, router, hunter, verifier, all reading from a tasks.json. It was the right shape for bug-hunting. For UI audits, it isn't enough. Reading Hermes this week, I saw the ceiling: a platform where taste compounds across sessions.
Fixing AI Slop in the 80%
Pareto applies to code: 20% is skeleton, 80% is garbage. Backend garbage is tractable — you can verify it. UI garbage is where AI slop lives, because taste isn't testable. One concrete move against it: stop screenshotting competitors. Record them.
I Sent 7 AI Agents to Fix Bugs Across 3 Repos. Here's What Broke.
AI agents can write code. But they can't think about thinking. I built an orchestration system to push them further — and found the ceiling.
Building an LLM Agent to Play Dwarf Fortress
Not because it's hard. Because it's abstract. How we wired an LLM to control Dwarf Fortress via DFHack — and got Phase 1 to actually run.
Clawdbot: Building My AI Companion from Scratch
How I set up a self-hosted AI agent that actually remembers me, checks my health data, and automates my boring tasks.
On TTA: The Art of Bottlenecks
Through the Ages (TTA) is not just a game; it's a cold, mathematical simulation of civilization-building that teaches you the art of identifying bottlenecks.
On Life: The Paradox of the System
I'm writing this to avoid my life becoming shit
On Game
my game experience started from...
Recent Posts
On Self-Improvement：怎么知道它真的变好了
让 agent 改进自己的 harness，几乎总是在过拟合。写补丁很便宜，证明它变好了才是难的。有效的方法下面都有 benchmark，而我们只有一个收件箱。
On Markov: The Best Move From Here
My life is one giant Markov chain: only the current state matters. Except a chain has no player. Games hand you the win condition and a reload button. Life gives you one run and makes you write the reward yourself.
On Harness：从 loop 到 gate
六篇文章讲的是同一件事: 把自己从 loop 里挪到 gate 上. 以及为什么其中三篇自己没被打过分.
On Layers
Five pieces that survived every gear purge — including almost ten Echos, worn all four seasons. The closet keeps converging on OR, and not out of loyalty. Good gear is the gear you stop noticing.