07-11 14:38 · 开源情报,行业分析,趋势监控,决策支持,AI工具
Evidence-backed AI industry intelligence — trends, source updates, daily data refreshes, and weekly decision briefs.
English · 简体中文
Evidence-backed AI industry intelligence for people who need to decide, not just keep up. Agent Pulse turns official releases, research papers, capital moves, policy changes, and public propagation signals into traceable Events, evolving industry judgments, and explicit next signals.
Open Agent Pulse · ⭐ Star this repository · Read in Chinese
AI has no shortage of news, feeds, or hot takes. What is scarce is a reliable way for investors, executives, founders, business leaders, and technical leaders to answer three questions:
- What materially changed? Not every announcement deserves a new view.
- Why does it matter? A fact becomes useful only when connected to an industry shift and the people it affects.
- What would change the judgment? Every durable view needs a next signal, counter-signal, or invalidation condition.
Agent Pulse is built around that loop: observe broadly, publish selectively, preserve the evidence trail, and update a judgment only when the evidence changes.
If that is the kind of public intelligence infrastructure you want to exist, star Agent Pulse. It helps more decision-makers and builders discover the project and is the clearest signal that the open version is worth maintaining.
Every public event is designed to separate six layers:
- Fact — what happened and when;
- Evidence — the original release, paper, filing, or policy document;
- Context — what came before and what actually changed;
- Impact — who is affected across technology, business, capital, and policy;
- Judgment — the current interpretation, with uncertainty made explicit;
- Next signal — what would strengthen, weaken, or invalidate the judgment.
official releases + papers + filings + expert and propagation signals
│
▼
collect → normalize → deduplicate → cluster → evidence binding
│
▼
deterministic gate → strategic narrative → next signal → public event
Aggregators can suggest candidates or propagation heat, but cannot become the sole factual evidence for a material event. External text is never treated as trusted HTML, and private operational data never enters the public static export.
The repository is not a mockup. It contains the source catalog, collectors, evidence model, automatic quality gates, static renderer, source-health automation, and GitHub Pages delivery path used by the live product. It also keeps four states deliberately separate: catalogued, observed, active, and published.
GitHub Actions refreshes public data and redeploys the static site once per day. The source audit, monitor, and quality guard remain weekly; the weekly-brief
Issue is created or updated only on Sunday (or by an explicit manual run) and only when at least one public Event clears the weekly gate, so daily freshness does not create empty Issues.
Repository evidence checked on 2026-07-14. Source-health numbers below come from the full audit completed at 2026-07-14 05:36 UTC; content counts come from the versioned repository snapshot.
See the machine-generated source health report, system capability evaluation, data-source policy, and capability map. CI restores the complete repository snapshot, compares a fresh evaluation with the previous commit, fails on score or evidence regression, and publishes the highest-priority evidence gaps as its improvement queue.
The public source-update stream exposes only allowlisted titles, attribution, dates, categories, tags, and canonical links. It remains an observation layer and cannot bypass the evidence and readiness gates required for Event publication. The limitations are equally important: production qualification still needs a real observation window; many historical events need more independent evidence; claim-level evidence, multilingual semantic clustering, real MySQL integration coverage, and user outcome feedback are still incomplete. Planned or experimental capabilities are never presented as shipped.
sources and discovery signals
│
SourceAdapter boundary
│
safe fetch → normalize → quality gate → deduplicate
│
isolated observation / event clustering
│
evidence binding → deterministic readiness gate
│
┌────┴────┐
▼ ▼
Control Room privacy-safe static site → GitHub Pages
SQLite is the zero-configuration default. A MySQL dialect path exists, but the project does not claim MySQL compatibility without real integration coverage. Public Pages contain only allowlisted DTOs; databases, credentials, raw payloads, proxy settings, and private notes are excluded.
Read the architecture, roadmap, or changelog for implementation details and release history.
Requires Node.js 22 or later.
git clone https://github.com/barretlee/agent-pulse.git
cd agent-pulse
npm install
cp .env.example .env
npm run dev
Local startup automatically migrates the database, refreshes catalog seed metadata, and merges the latest versioned repository snapshot. A fresh clone and an existing local SQLite database therefore start from the same complete repository dataset without deleting newer local evidence. npm run db:
▸ 展开全文