# ELITE Agentic Engineers Are LEAVING THE LOOP... Here’s why

## Executive summary

The video outlines the critical evolution of agentic engineering, arguing that human engineers are becoming the bottleneck in the development cycle. The focus must shift from 'Human In The Loop' (HITL) prompting to building 'Near Loop' and ultimately 'Out Loop' systems. The next leap requires building proactive, durable agents that operate autonomously in isolated sandboxes, triggered by events (e.g., webhooks, log streams) rather than direct human attention. Mastery of the 'Core Four'—Context, Model, Prompt, and Tools (Harness Engineering)—is essential for building these highly leveraged, specialized systems.

## Key takeaways

- The Agentic Transition: In, Near, to Out the Loop: The goal is to move from HITL (where attention is constantly applied to prompts) to Near Loop (where the human provides approval/clicks on a concrete decision, e.g., a patch) to Out Loop (full cron autonomy, where the system operates without human intervention).
- The Bottleneck Problem: The speaker asserts that up to 80% of current engineering work can run without human input if the correct system architecture is implemented, making the engineer the primary limitation.
- Core Architecture for Proactivity: To achieve proactive agents, systems must combine Durable Agents (which survive crashes/restarts) with isolated sandboxes (providing dedicated CPU/GPU compute) and a clear API layer for varied triggers (Cron, Webhooks, Log Events, Jev classification).
- The Core Four of Agentic Engineering: The four essential skills for advanced agents are Context, Model, Prompt, and Tools (Harness Engineering). Specialization—building systems that outperform default agents—is the key advantage.

## Technical details

- Agent Capabilities & Models: Advanced models like Opus 5.5 and GPT 6 Astra are claimed to be superior programmers to humans, capable of complex, multi-step tool calls, necessitating a shift in operational strategy.
- Near-Loop Workflow Example (Incident Response): A near-loop system can detect an event from a log stream, use Jev to classify the event, run root-cause analysis, prepare a fix, and then wait for a single human click for deployment, minimizing manual effort.
- Durable Agents: Durable agents are required because they persist, survive restarts, and maintain memory, unlike classic in-the-loop agents, enabling continuous, background work.
- System Architecture: The ideal setup involves running agents in a sandbox environment, connected via an API layer, allowing for triggers like Cron, log events, or user actions, thereby enabling proactive operation.

## Practical implications

- Focus development efforts on building system architectures (the system that builds the system) rather than optimizing individual prompts or agents.
- Implement event-driven triggers (webhooks, log streams) to move agents from reactive prompting to proactive, autonomous operation.
- Prioritize building agents within isolated, durable sandboxes to ensure persistence and reliable, long-running compute.
- Master the 'Core Four' skills (Context, Model, Prompt, Tools) to specialize agents and achieve unique, high-leverage results.

## Topics

Agentic Engineering, Build Automation, Durable Agents, Software Architecture, AI Workflow, CI/CD, IA3 Phase 3 Early Sign Up, Proactive Durable News Agent, Proactive On-Call Durable Agent

Source: https://www.youtube.com/watch?v=52JX83rGPLc