wuyang
69 lines
1.8 KiB
Markdown
69 lines
1.8 KiB
Markdown
# Industry: Microsoft Research - AgentRx
|
|
|
|
---
|
|
type: industry
|
|
company: Microsoft
|
|
team: Microsoft Research
|
|
title: "Systematic debugging for AI agents: Introducing the AgentRx framework"
|
|
url: https://www.microsoft.com/en-us/research/blog/systematic-debugging-for-ai-agents-introducing-the-agentrx-framework/
|
|
source_name: Microsoft Research
|
|
source_type: research-blog
|
|
source_quality: official
|
|
published_at:
|
|
collected_at: 2026-07-08
|
|
status: analyzed
|
|
topics:
|
|
- agent-evaluation
|
|
- observability
|
|
- debugging
|
|
implementation_signals:
|
|
- trajectory-normalization
|
|
- constraint-synthesis
|
|
- constraint-checking
|
|
- failure-taxonomy
|
|
product_area:
|
|
- agent-debugging
|
|
models:
|
|
-
|
|
tools:
|
|
- AgentRx
|
|
benchmarks:
|
|
- AgentRx Benchmark
|
|
related_papers:
|
|
- 2025-vijayvargiya-openagentsafety
|
|
related_jobs:
|
|
- 2026-07-08-tencent-cloud-ai-agent-test-engineer
|
|
related_experiments:
|
|
-
|
|
related_projects:
|
|
-
|
|
evidence_level: high
|
|
relevance: high
|
|
---
|
|
|
|
## One-line Takeaway
|
|
|
|
Agent 调试需要从“最终失败”定位到 trajectory 中第一个不可恢复错误步骤。
|
|
|
|
## What They Built or Claimed
|
|
|
|
Microsoft Research 发布 AgentRx 和 115 条人工标注失败轨迹 benchmark，用于定位 Agent trajectory 的 critical failure step。
|
|
|
|
## Technical Signals
|
|
|
|
- architecture: trajectory normalization -> constraint synthesis -> guarded evaluation -> LLM judge。
|
|
- tool use: 根据 tool schema 和 domain policy 生成可执行约束。
|
|
- memory:
|
|
- evaluation: 逐步检查约束，输出可审计验证日志。
|
|
- safety: 区分计划偏离、幻觉、新信息发明等失败类型。
|
|
- deployment: domain-agnostic debugging framework。
|
|
- data: tau-bench、Flash、Magentic-One 失败轨迹。
|
|
|
|
## Evidence Quality
|
|
|
|
官方研究博客，包含框架、benchmark 规模、失败分类，证据强。
|
|
|
|
## Gaps for Us
|
|
|
|
- 本知识库需要设计统一 Agent trace schema 和失败 taxonomy。
|