Run of Tuesday, October 27, 2026 · real pipeline run

How it works

Five LLM agents do the reading; plain code does the counting and the deciding. Each box shows how many items went in and came out in this run. Click one to see a real input and output.

30 loaded → 11 kept → 49 shifts → 38 published → 114 skill mentions → 14 in a quadrant → 3 brief items

News sideBoth sides meet

2. Filter

src/agents/filter.ts · prompt: src/agents/prompts/filter.ts

Scores each article's relevance to hiring for the role and spots duplicates. Code makes the keep or drop call from a config threshold.

This run: 3 dropped as off topic or duplicate.

View codeFilter agentL33-59 PromptL5-36 Orchestrator stepL82-94

LLM wrapperValidation + one retryL89-127 validate()L143-157

14 calls25,725 input tokens1,412 output tokens32.6 s summed

Sample input

[
  {
    "id": "news-012",
    "title": "Making Your Data Ready for Agentic AI",
    "source": "martinfowler.com",
    "url": "https://martinfowler.com/articles/making-data-ready-for-agentic-ai.html",
    "publishedDate": "2026-08-27",
    "sourceType": "blog",
    "placeholder": false
  },
  {
    "id": "news-007",
    "title": "Have a Meta business account? You're consenting to AI scraping now, legal experts say",
    "source": "ZDNET",
    "url": "https://www.zdnet.com/business/meta-business-account-data-ai-scraping-terms/",
    "publishedDate": "2026-10-16",
    "sourceType": "trade",
    "placeholder": false
  }
]

Sample output AI-generated · Filter

[
  {
    "articleId": "news-012",
    "keep": true,
    "relevance": 0.78,
    "reason": "Describes a concrete shift in data practice for agentic AI (data contracts, traceability, governance, context and access layers) that TPMs running AI and data programs will increasingly be asked to plan and oversee, though it is architectural guidance rather than a hiring announcement.",
    "duplicateOf": null
  },
  {
    "articleId": "news-007",
    "keep": false,
    "relevance": 0.3,
    "reason": "This is a publisher and platform terms-of-service dispute over AI scraping of Meta business accounts, with no clear change to the skills or practices TPM employers expect.",
    "duplicateOf": null
  }
]

How accurate are the agents?

Shift Extraction11 articles, hand-labeled

Precision: 96%
Recall: 100%

JD Analyzer5 postings, hand-labeled

Precision: 98%
Recall: 100%

Required vs preferred right 100% of the time

Compared per (item, skill) pair. Precision: of the skills the agent found, how many a human also labeled. Recall: of the skills a human labeled, how many the agent found.

Where the agents and the labels disagree (3)

news-004: extra api-platform-integration
news-002: extra agile-program-execution
jd-tpm-009: extra agentic-workflows

Cost and latency log

LLM calls: 53
Input tokens: 120,404
Output tokens: 23,660
Agent time: 221.9 s

64,966 input tokens were read from the prompt cache. Agent time is the sum of every call; calls run in parallel, so the run itself is shorter. Models: claude-sonnet-5-5.

Agent	Calls	Input	Output	Cache read	Time	Retries
filter	14	25,725	1,412	0	32.6 s	0
shift-extraction	11	20,301	8,137	26,536	66.1 s	1
verification	11	46,002	5,256	0	50.6 s	0
jd-analyzer	16	25,362	7,480	38,430	60.9 s	2
brief-writer	1	3,014	1,375	0	11.6 s	0