My Jev Pi Durable Agent CURATES News: HYPE, FACT, or SIGNAL

Summary

Indie Dev Dan demonstrates a persistent, proactive news-curation agent designed to reduce the time engineers spend sorting through AI-related news. Instead of waiting for a prompt, it runs on a schedule, gathers and ranks items, then waits for user feedback.

Core idea: hype, facts, and signal

The agent’s curation is guided by three categories:

• Hype: Attention driven mainly by excitement or promotion.
• Facts: Claims that are objectively true but may not be useful to act on.
• Signal: Information that is both relevant and actionable for the user.

For example, the agent might classify a new model release as a useful signal while also noting that its announcement contains some hype.

Architecture and workflow

The system is presented as a client-server application:

• A persistent agent runs behind an API and is backed by SQLite.
• A cron trigger starts scans about once per minute.
• Browser, terminal/TUI, and CLI interfaces act as lightweight clients connected to the same agent, keeping their views synchronized.
• The agent can run on an M4 Mac mini or in a cloud sandbox. The demonstration shows separate instances with different feeds and preferences.

During a scan, the system collects news, filters and deduplicates items, researches claims across sources, and ranks candidates for presentation. The speaker says Jev is used for classification, relevance checks, duplicate detection, and ranking, helping reserve more expensive language-model calls for promising items. Each run can surface up to seven cards.

Product features and feedback

The feed is personalized through a data flywheel: user actions help tune future results.

Users can:

• Like or dislike stories to adjust their ranking.
• Ask the agent to change keyword weights; in the demo, the weight for “Gemini” is increased.
• Customize preferred keywords and news sources.
• Open source links to inspect the underlying reporting.
• Review process traces and past executions.

The agent has purpose-built tools for news review, candidate lists, item reading and research, web search, URL scanning, classification, timelines, and preference management. The speaker emphasizes that these product-specific actions and the system design matter more than merely generating code.

Main takeaway

The video argues that agent engineering should progress from working “in the loop,” to “near the loop,” and eventually “out of the loop.” Persistent agents and scheduled triggers are presented as a practical way to automate recurring work while retaining human oversight through feedback and customization. The speaker also stresses managing token costs and treating agents as tools that still require engineering judgment.

Guide and references mentioned

• A walkthrough of the news agent’s architecture, workflow, and user-facing features.
• A reference to a previous video demonstrating a Pi Durable incident-response agent.
• The speaker points viewers to the news-agent codebase and promotes a forthcoming Phase 3 agent-engineering course.

Main speaker: Indie Dev Dan

Technologies, projects, and sources mentioned: Pi Durable, SQLite, Jev and the Type Safe team, cron, API, browser/TUI/CLI clients, M4 Mac mini, exe.dev sandbox, and news sources including Hacker News, Latent Space, Simon Willison, and AI-company announcements.