Skip to content

Instantly share code, notes, and snippets.

View weiqiangt's full-sized avatar
:electron:

Weiqiang Tang weiqiangt

:electron:
  • Beijing, China
  • 20:30 (UTC -12:00)
View GitHub Profile

SAGE 模型:LLM Agent 工具体系的四象限分类与演进趋势

撰写时间:2026 年 3 月 | 读者在引用本文结论时,请对照自身所处时间点,校准预测的时效性。

在大语言模型(LLM)驱动的 Agent 系统中,工具(Tool)与技能(Skill)的设计核心在于对冲 LLM 的架构缺陷与概率不确定性。将一个功能封装为 Tool 的判定标尺是:该任务 LLM 理论上能做,但大规模场景下极易出错、消耗极大算力(Token),或缺乏与外部真实世界的连接。

本文提出 SAGE 模型,将 Agent 的工具箱从通用软件工程与任务执行的视角划分为四个象限——Solve(求解)、Anchor(锚定)、Gate(门控)、Expert(专鉴),并对每个象限给出短期(1 年内)、中期(2–3 年)和长期(5 年以上)的演进预测。

SAGE 象限定位图

@weiqiangt
weiqiangt / Profile Rust on Linux.md
Created December 22, 2018 08:30 — forked from KodrAus/Profile Rust on Linux.md
Profiling Rust Applications

Profiling performance

Using perf:

$ perf record -g binary
$ perf script | stackcollapse-perf.pl | rust-unmangle | flamegraph.pl > flame.svg

This requires you have flamegraph available in your path. The rust-unmangle script is optional but nice.

@weiqiangt
weiqiangt / ObjectHeader32.txt
Created September 18, 2018 11:50 — forked from arturmkrtchyan/ObjectHeader32.txt
Java Object Header
|----------------------------------------------------------------------------------------|--------------------|
| Object Header (64 bits) | State |
|-------------------------------------------------------|--------------------------------|--------------------|
| Mark Word (32 bits) | Klass Word (32 bits) | |
|-------------------------------------------------------|--------------------------------|--------------------|
| identity_hashcode:25 | age:4 | biased_lock:1 | lock:2 | OOP to metadata object | Normal |
|-------------------------------------------------------|--------------------------------|--------------------|
| thread:23 | epoch:2 | age:4 | biased_lock:1 | lock:2 | OOP to metadata object | Biased |
|-------------------------------------------------------|--------------------------------|--------------------|
|
@weiqiangt
weiqiangt / samm
Created June 5, 2015 13:25
SharedMem Accelerate Matrix Multiply
__global__ void MatrixMultiKernel(float *d_M, float *d_N, float *d_P, int width){
__shared__ float M_d_s[TILE_WIDTH][TILE_WIDTH];
__shared__ float N_d_s[TILE_WIDTH][TILE_WIDTH];
int bx = blockIdx.x, by = blockIdx.y;
int tx = threadIdx.x, ty = threadIdx.y;
int Row = by*TILE_WIDTH + ty;
int Col = bx*TILE_WIDTH + tx;