wuyang
3.8 KiB
3.8 KiB
Industry Insights
status: seed
这个文件沉淀大公司 Agent 工作和技术报告的趋势判断。单条资料放在 industry/items/，这里记录跨公司的归纳。
Current Snapshot
- date: 2026-07-08
- scope: large-company Agent work, engineering blogs, benchmarks, product/technical reports
- sources reviewed: 9 starter items
- source mix: official 9
Strong Industry Signals
- Enterprise Agent 正在从 demo 转向有权限、记忆、评估、trace、安全和平台入口的系统。
- Computer use/GUI Agent 正在成为主流产品能力，但安全控制必须产品化。
- Memory 的工业定义正在从“记住信息”转向“通过经验改善流程执行”。
- Agent debug/eval 正在围绕 trajectory、constraint、failure taxonomy 和 benchmark 构建。
- 大公司开始把 Agent adoption 当成组织生产力指标看，而不是单一模型能力。
Weak or Marketing-heavy Signals
- 单纯模型发布里提到 “agent frontier” 但缺少 benchmark、代码或系统细节时，只能作为方向信号。
- Qwen-AgentWorld 当前抓取到的信息不足，需要补完整正文。
- Agent usage growth 指标能说明采用趋势，但不能直接说明任务质量。
Company Patterns
| Company | Agent Direction | Evidence | Notes |
|---|---|---|---|
| OpenAI | internal data agent, enterprise adoption, Codex as work agent | in-house data agent, agents transforming work | 系统细节集中在数据 Agent |
| Anthropic | computer use and coding agents | not yet itemized | 下一轮补 Claude computer use / Claude Code |
| Google DeepMind | computer use, agent security/control | Gemini computer use, AI Control Roadmap | 安全控制非常值得沉淀 |
| Microsoft | memory benchmark, failure diagnosis, skill optimization | STATE-Bench, AgentRx, SkillOpt | 评估和工程方法密集 |
| DeepSeek | Agent integration ecosystem and hiring wave | jobs item, awesome-deepseek-agent source candidate | 需补官方技术资料 |
| 腾讯 | Agent testing / cloud harness signal | jobs item | 需补技术博客和北京岗位 |
| 字节 Seed | productivity agent, GUI+MCP, coding agent | Seed2.1 official blog | 与岗位中的 harness/memory/context compression 呼应 |
| 阿里 Qwen | AgentWorld / world models for general agents | Qwen-AgentWorld queued item | 需补正文和 Qwen3.x agent blogs |
Implementation Patterns
- tool use: API/MCP/function calling 与 GUI/computer use 两条线并行。
- memory: 从 raw history 转向 structured memory、stateful benchmark、experience consolidation。
- evaluation: adoption metrics、benchmark、trajectory failure diagnosis、deterministic assertions。
- safety: defense-in-depth、sandboxing、human confirmation、prompt injection guard、permission gating。
- deployment: Slack/Web/IDE/CLI/MCP、enterprise platform、internal tools。
- productization: coding agent、data agent、office/productivity agent、customer support/travel/shopping workflows。
Follow-up Queue
- Anthropic Claude computer use / Claude Code / model cards。
- OpenAI workspace agents / Frontier / Codex enterprise coding posts。
- Qwen AgentWorld and Qwen3.x agent blogs full extraction。
- DeepSeek official GitHub agent ecosystem and talent site details。
- Tencent AI Lab / Hunyuan / Yuanbao Agent technical posts。
Links Back to Knowledge Base
- docs to update: evaluation, observability, memory, safety, practice playbook
- papers to read: OpenAgentSafety, memory survey, EvoMemBench, SWE-EVO, Dialogue-SWEBench
- jobs to compare: Baidu Agent algorithm/fullstack, DeepSeek Agent hiring, Tencent Agent testing, ByteDance Seed Agent research
- experiments to create: memory benchmark, trace/failure diagnosis, GUI+MCP workflow
- projects affected: future knowledge frontend, coding-agent experiments, eval harness