How to Run Loop Engineering with Hermes and Claude

TL;DR

Loop engineering shifts the focus from writing prompts to designing self running systems that drive the agent. It defines five parts of a loop, explains two loop types, and shows how long running tasks become possible with Herm es and Claude. The video explains how to set up autonomous loops for persistent projects and reduce manual prompting.

Transcript

There's a new term going around and you might have already heard it. It's called loop engineering. And just like every other hype term, everyone is talking about it like it's something new. It's not. But when you combine it with an always running agent like Hermes, it stops being hype. Most people who are trying to set these up are getting the loop... Read More

Key Insights

• Loop engineering shifts control from prompts to autonomous system design.
• There are two loop types, deterministic and non deterministic, each with different use cases.
• Five core loop components are context management, feedback quality, verification gates, termination conditions, and state management.
• Context management prevents important information from being buried by recent outputs.
• Feedback quality is a critical signal guiding the next action for the agent.
• Verification gates convert feedback into a clear done status for the loop.
• Termination conditions must be explicitly defined to avoid premature or endless looping.
• State management across turns relies on external files to keep track of progress beyond the context window.
• More videos from this channel:
  • How to use GitHub Spec Kit for AI coding workflows?
  • How to Prevent Context Rot in Claude Code
  • What is MCP in AI tools and how does agents.md help?
• More from this channel: AI LABS

Questions & Answers

Q: What is loop engineering and how does it differ from prompt engineering?

Loop engineering is the practice of designing systems that generate prompts for the agent and drive the agent to complete tasks autonomously. Unlike prompt engineering, which concentrates on crafting prompts, loop engineering handles the entire cycle from state assessment to action, feedback, verification, and completion. This shift reduces the need for manual prompting and aims to keep the agent progressing toward the defined end goal.

Q: What are the two main loop types discussed and when should each be used?

The two loop types are deterministic loops and non deterministic loops. Deterministic loops are suited for tasks with a clear definition of done, such as tests passing or code compiling. Non deterministic loops are used for tasks without a strict finish rule, like building a user interface, where progress must be evaluated by an external verifier and may involve ongoing refinement.

Q: What are the five steps of building a loop as described in the video?

The five steps are: 1) check the project state to understand current conditions, 2) let the model decide the next action, 3) execute the action by calling tools and running commands, 4) gather feedback from outcomes, and 5) decide if the task is done or continue. These steps are part of both loop types but adapt slightly based on the loop design.

Q: Why is context management important in loop engineering?

Context management matters because the agent cannot rely on the entire chat history to maintain useful information. As the task evolves, important details can be buried under new outputs, even with large context windows. Proper context management ensures the agent retains the critical information necessary to progress toward the end goal.

Q: What role does feedback quality play in loop success?

Feedback quality is a key signal that tells the agent how it performed. It can take forms such as test results or UI screenshots, and the agent uses this information to inform its next move. High quality feedback helps steer the loop toward correct behavior and completion.

Q: What are verification gates in the loop, and why are they needed?

Verification gates are checkpoints that translate feedback into a clear verdict about whether the task is finished. They prevent premature completion or endless looping by providing concrete criteria that must be met before the agent marks the task as done.

Q: What is the purpose of a termination condition in a loop?

The termination condition defines when the loop should stop. It must be explicit to avoid stopping too early or continuing without real progress. Without a clear termination rule, the loop can drift away from the intended end goal.

Q: How does the video describe managing state across turns, and why is it necessary?

State management across turns involves keeping track of task progress with external files because the context window cannot hold everything forever. This approach lets the agent continue working without losing the thread, ensuring continuity even as the conversation grows and context becomes less reliable.

Summary & Key Takeaways

• Loop engineering replaces prompt writing with designing autonomous loops that guide agents, enabling sustained tasks. The video compares loop engineering to traditional prompt engineering and explains why autonomous loops perform better as projects scale. It also highlights the roles of context management, feedback, verification gates, termination conditions, and cross turn state.

• Two loop types are explored: deterministic loops for defined done criteria and non deterministic loops for open ended tasks. The deterministic approach is shown with examples like tests and deployments, while non deterministic loops are used for UI tasks using an AI Slop Detector and a verifier model.

• The five step flow to build loops is introduced, followed by a look at how to implement loops with Claude Code and Hermes. The discussion emphasizes long term project readiness, model capability, and the need to structure systems so the agent can operate with minimal human intervention.