Miadi Review

Engineering Proactive and Durable Out-of-the-Loop Agents

Oct 25, 2026 · model: gemini-3.7-flash · channel: IndyDevDan

Watch source

Engineering Proactive and Durable Out-of-the-Loop Agents

Source: My Jev Pi Durable Agent CURATES News: HYPE, FACT, or SIGNAL — IndyDevDan https://youtu.be/OLqP8jxJ0Wc

Summary

IndyDevDan demonstrates how modern software engineers can transcend manual prompting by building proactive, durable agents that migrate workflows from human-in-the-loop to near-the-loop and out-of-the-loop. Implementing a continuous news curation engine called Durable News Agent, he pairs an M4 Mac Mini server, SQLite persistence, and Pi Durable with JEV-based cost gating to automatically sweep, deduplicate, and score tech news into hype, facts, and signal. Decoupled web and terminal clients communicate over a unified SDK stream, empowering engineers to direct agent priorities through dynamic weight tuning and two-way prompting while autonomous background loops execute continuously.

Aspects

In-Loop to Out-Loop Progression (00:35-01:17)

Dan frames engineer leverage through transitioning from human-in-the-loop to human-near-the-loop, and finally out-of-the-loop. Rather than waiting for manual user prompts, proactive agents run autonomously on scheduled cron intervals, allowing engineers to inspect generated feeds, provide feedback, and step away while systems continue working independently.

Hype, Facts, and Signal Triad (05:15-06:33)

Navigating the AI deluge requires categorizing incoming data into hype, facts, and actionable signal. Dan emphasizes engineering taste and judgment: distinguishing objectively true facts from operational signal developers can immediately act on, which guides how agents score, filter, and surface relevant intelligence.

Detached Clients and Shared State (02:00-02:45)

The agent daemon runs remotely on an M4 Mac Mini backed by SQLite databases, exposing an API rather than binding to a single UI. Detached clients—web browsers, terminal interfaces (TUI), and CLIs—connect to a single real-time stream via the DNA SDK, maintaining synchronized state across multiple devices.

High-Throughput Cost Gating with JEV (03:54-04:15)

Running recurring background sweeps threatens token budgets if routed purely through frontier LLMs. Dan inserts JEV as a cheap, ultra-fast classifier stage costing fractions of a cent per call to filter, deduplicate, and gate candidate links before triggering downstream deeper LLM synthesis.

Layered Agentic Software Architecture (08:30-09:08)

Durable autonomous workflows demand modular layered architecture: persistent storage (SQLite), runtime execution harnesses (Pi Durable), domain API logic (news-desk), lightweight classification (JEV), and client SDK distribution (DNA SDK). Separating concerns ensures the agent remains an adaptable, reproducible platform rather than a brittle script.

End-results and Core Aspect

Core Aspect — Continuous Autonomous News Curation Engine (01:35-02:55)

A durable, proactive agent deployed on an M4 Mac Mini continuously harvests, gates, classifies, and ranks technical feeds every minute. By operating completely detached via cron and local SQLite checkpoints, it transitions the engineer from manual curation to near-the-loop supervision with zero downtime.

Results

  • Multi-Client Synchronized Feed Inspection (02:15-02:40): Changes made via terminal CLI or generative web UI instantly synchronize through the DNA SDK stream, letting the developer review ranked cards, inspect sweep executions, and audit system traces from any detached client without interrupting ongoing cron sweeps.
  • Cost-Effective Autonomous Link Sweeps (03:50-04:15): The engine executes dozens of automated daily sweeps across developer news sources for approximately $0.60 to $1.90 per day. Chaining JEV classification ahead of GPT-6 Luna or frontier models drastically cuts token compute while keeping fresh breaking releases surfaced.
  • Real-Time Two-Way Preference Steering (13:48-14:48): The engineer steers curation quality live by upvoting or downvoting cards and submitting natural-language instructions into a bidirectional prompt system. The agent dynamically adjusts SQLite-stored topic weights, immediately altering prioritization thresholds for subsequent background sweeps.
  • Multi-Environment Sandboxed Deployment (07:35-08:06): The identical codebase deploys interchangeably across local hardware (M4 Mac Mini) and cloud sandboxes (exe.dev), running independent specialized news desks that prove architecture portability and isolation across local and hosted execution runtimes.