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

## Executive summary

The video details the concept of 'proactive agents'—a critical evolution in agentic engineering that moves systems from 'in the loop' to 'near the loop' to 'out the loop.' The speaker demonstrates a Durable News Agent, which runs autonomously (e.g., via cron intervals) on an M4 Mac mini, continuously gathering, filtering, and curating high-signal engineering news. The architecture emphasizes building robust, decoupled client-server systems using technologies like Pi Durable and SQLite, allowing the agent to operate without constant human intervention.

## Key takeaways

- The Agentic Engineering Shift (In, Near, Out): The core meta-lesson is the transition from 'human in the loop' (manual prompting) to 'human near the loop' (supervising agents) to 'human out the loop' (autonomous systems). This shift represents a measure of advanced agentic engineering skill, allowing systems to work continuously for the user [0:01:12].
- Proactive Agent Architecture: A proactive agent is defined as a durable agent combined with a trigger inside a sandbox. The system operates on a client-server model, where the browser UI and terminal TUI are detached clients synced to the core agent running on the M4 Mac mini [0:01:27].
- Information Filtering (Hype, Fact, Signal): The speaker proposes a framework for filtering AI noise: 'Hype' (unfounded excitement), 'Fact' (objectively true but not actionable), and 'Signal' (actionable, high-value information). This judgment is crucial for building useful systems [0:02:15].

## Technical details

- System Architecture: The Durable News Agent uses a layered, client-server architecture: 1) SQLite databases for persistence; 2) Pi Durable for durable tasks; 3) A custom Durable Agent exposed via an API; 4) Jev for classification, filtering, and ranking; and 5) an SDK used by the browser, TUI, and cron jobs [0:03:40].
- Data Processing Pipeline: The workflow involves a top-level news sweep, followed by tasks that gather, gate, deduplicate, and rank candidates. Jev is used at multiple points to classify large amounts of text at sub-half-second latency, saving compute and token costs [0:03:56].
- Feedback Loop Implementation: The system incorporates a data flywheel by allowing the user to upvote or downvote news items (e.g., Gemini 4 Argon release). This user feedback is passed back to the agent to increase or decrease the weight/signal of specific keywords (e.g., increasing the weight of 'Gemini' from 1.2 to 1.3) [0:04:40].

## Practical implications

- Build systems that operate proactively and autonomously (out-of-loop) to maximize leverage and minimize manual intervention.
- Implement a structured data flywheel by integrating user feedback (up/downvoting) directly into the agent's ranking and weighting algorithms.
- Utilize lightweight classifiers (like Jev) early in the pipeline to gate expensive LLM calls, optimizing token usage and cost.
- Focus on unique system architecture and judgment (taste) rather than just code generation, as code is becoming infinitely cheap.

## Topics

Agentic Engineering, Proactive Agents, Durable Computing, System Architecture, AI Workflow, Durable News Agent codebase, Pi Durable video

Source: https://www.youtube.com/watch?v=OLqP8jxJ0Wc