The Creators of Claude Code and OpenClaw don't Prompt Their Agents Anymore?!

Key Concepts

• Loop Engineering: A methodology where AI coding agents are configured to run in iterative cycles (loops) to perform tasks autonomously, rather than relying on single-turn prompts.
• Orchestrator/Worker Pattern: A system architecture where a primary "orchestrator" agent manages high-level goals and dispatches tasks to specialized "worker" agents.
• Harness Engineering: The practice of building deterministic wrappers or frameworks around AI agents to manage state, cost, and reliability.
• Human-in-the-Loop (HITL): A design principle requiring human intervention or approval at specific stages of an automated workflow.
• Durability: The ability of an AI workflow to persist state in an external database (e.g., PostgreSQL/Neon) so that tasks can be resumed after failures or interruptions.
• Context Bloat: The phenomenon where long-running agent sessions accumulate excessive data, leading to performance degradation and increased token costs.

---

1. The Concept of Loop Engineering

Loop engineering is currently being popularized by figures like Peter Steinberger (Open Claw) and Boris Cherney (Claude Code). The core premise is to move away from manual prompting toward designing autonomous loops that handle tasks 24/7.

• Core Mechanisms in Claude Code:
  • /loop: Sets an interval for an agent to wake up and check external systems (e.g., GitHub issues).
  • /goal: Defines success criteria, forcing the agent to work until the objective is met.
  • /routines: Scheduled jobs that trigger the agent to process larger documents or tasks incrementally.

2. Critical Trade-offs and Challenges

The speaker argues that "loop engineering" is often over-hyped and presents significant practical risks:

• Reliability: Autonomous loops often produce low-quality results if not strictly controlled.
• Token Consumption: These systems are "token hungry" because the orchestrator must constantly reason about state, manage workers, and pass context back and forth.
• Context Bloat: Running loops within a single session overwhelms the LLM’s context window.
• Scalability: The claim that one can manage "tens of thousands of agents" is questioned as potentially impractical and prone to bugs.

3. Methodologies for Robust Automation

To solve the issues of cost and reliability, the speaker advocates for Harness Engineering—building deterministic workflows that limit the agent's decision-making power.

• Deterministic Workflows: Instead of letting the agent drive the entire process, use a workflow file to define the steps (e.g., Extract Issue -> Classify -> Research -> Implement -> Validate).
• Distributed Sessions: Rather than one long session, each step in a workflow should run in its own isolated coding agent session.
• Cost Optimization: Use different models for different nodes. For example, use smaller, cheaper models (e.g., Haiku, Kimmy K2.7) for classification and context loading, and reserve more powerful models (e.g., Claude Opus) only for complex implementation tasks.
• State Management: Use external databases (e.g., Neon/Postgres) to store logs and progress, ensuring the system is "durable" and can resume from the exact point of failure.

4. Real-World Application: The Dashboard Approach

The speaker demonstrates an open-source dashboard designed to manage autonomous tasks effectively:

• Observability: The dashboard provides a visual interface to monitor the orchestrator’s decisions and worker status.
• Human-in-the-Loop Integration: The system pauses at critical junctures, allowing the user to validate progress before the orchestrator proceeds to the next round.
• Deployment: Using platforms like Retool, these local control systems can be deployed to the cloud, allowing for team collaboration, permission management, and audit trails.

5. Notable Quotes

• "The real skill is designing loops that prompt your agents so they work for you 24/7." (Attributed to the current trend in AI development).
• "I don't prompt Claude anymore. I write loops and the loops do the work." (Boris Cherney, creator of Claude Code).
• "I would just fold loop engineering into harness engineering. It doesn't quite deserve its own buzzword." (The speaker's perspective).

Synthesis and Conclusion

While loop engineering offers a powerful vision for autonomous AI agents, it is currently hindered by high costs and reliability issues. The path forward is not to let agents run wild, but to build harnesses—structured, deterministic workflows that utilize external state management, model-specific cost optimization, and human oversight. By treating these systems as software engineering projects rather than simple prompt-based tasks, developers can create reliable, scalable, and cost-effective AI-driven automation.