Codex vs AnySearch: Why Source Quality Beats Answer Fluency in AI Research

The author investigates Forward Deployed Engineer (FDE) roles using Codex with two different search backends: the default search and AnySearch MCP. The goal is to evaluate how the search tool affects research quality when the model and prompt remain constant.

Round 1: Trend Judgment

Prompt:

调研最近 30 天 FDE(Forward Deployed Engineer)的岗位变化,并回答:哪些 Al 公司正在招聘或扩充这类岗位;它与 Al Engineer、Solution Architect 有什么区别;企业为什么需要工程师进入真实业务现场,把模型能力变成可交付结果。
要求优先使用企业招聘页、官方博客、团队分享和可核验的岗位数据;相同招聘信息不要因为被多个网站转载而重复计算;对于FDE 正在爆发等判断,必须给出时间范围、样本口径和证据,不能只引用媒体标题。

Default search results: 23-company sample, unified timestamps with precise timezones, relatively rigorous scope. However, some citations came from job aggregator sites, secondary reposts, or articles where FDE appeared only in the title with irrelevant content. The answer looked complete but the supporting material lacked solidity.

AnySearch MCP results: Clear data priority design, more detailed and credible data. With sufficient trustworthy data, the model's conclusions became more accurate.

Key metrics comparison:

Primary source ratio: Default 95% (20/23 companies used official career pages or ATS); AnySearch 100% (17/17 quantified samples used official ATS, plus Google, Microsoft, Cursor as supplementary = 20/20).

Duplicate events: Both 0 after deduplication by official job ID (AnySearch: 293 final job records deduplicated by requisition/job ID).

Search/extract calls: Default — full call logs not retained; AnySearch — 21 effective calls: 1 search, 7 batch_search, 13 extract; 7 batch searches contained 35 queries. Including auth init and probes, 45 total tools/call requests.

Completion time: Default 26 min 16 sec; AnySearch 21 min 49 sec.

Task field completeness: Both 100% coverage of hiring changes, role differences, enterprise needs, and statistical scope. AnySearch additionally covered time window, sample scope, deduplication method, and "boom" judgment evidence.

Overall, AnySearch was slightly faster and provided finer-grained data.

Round 2: Real-World Tasks

Prompt:

FDE 究竟是一种新的工程岗位,还是 Solution Engineer、Al Engineer、技术顾问换了一个名字?
查找多家 Al 公司当前可访问的官方招聘页和团队分享,对比以下信息
- 日常工作是在写 Demo、做交付,还是维护生产系统
- 是否需要长期进入客户的业务现场;
- 工程能力、业务理解和沟通能力各占多大比重;
- 与 Al Engineer、Solution Architect、技术顾问的职责边界;
- 哪些要求是各家公司共有的,哪些只是个别公司的定义。
整理成一张FDE 岗位画像
- FDE 主要解决什么问题
- 日常工作包含哪些环节;
- 需要哪些核心能力;
- 更接近哪几类传统岗位
- 哪些人可能适合转向这个方向。

Default search: Took ~9 minutes, found only 9 companies. Conclusions were somewhat helpful but limited.

AnySearch MCP: Supports batch queries, completed in ~5 minutes. Cross-verified 10 companies' official career sites and related materials, yielding a detailed role research report with key conclusions. The 10-company comparison included granular breakdowns of engineering, business, and communication skill requirements.

Observation: AnySearch's advantage in vertical domains is more exhaustive data and higher execution fidelity.

Agent Search Must Optimize for Task Completion

The real value of AnySearch is not just an extra search entry point, but its ability to process results into evidence better suited for agent consumption. Key capabilities:

Routing search intent to match general or vertical sources.

Reducing probability of repeated same-source appearances.

Extracting full text and returning structured content, reducing agent re-cleaning effort.

Continuing remaining tasks when a local request fails.

Caveats: Two query sets only reflect performance in this environment, not all search tasks. Higher search quality ≠ guaranteed correct model judgment; important conclusions still need verification against original sources. Tool, MCP, Skill are merely capability entry points; task decomposition, prompt, model, and acceptance criteria equally affect outcomes.

Final insight: Search is not the endpoint; the true endpoint is enabling agents to produce results that can feed directly into articles, courses, projects, and knowledge bases. Readers using Codex, Claude Code, Cursor, or custom agents should test with a real task prone to duplicate/SEO noise to judge fit for their workflow.

AnySearch offers API, MCP, and Skill integration methods.