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

Too Long; Didn't Watch — Summary

Loop engineering is the practice of designing autonomous systems where AI agents prompt themselves in cycles to complete complex tasks, requiring robust harnesses to manage high costs and reliability issues.

Main Takeaways

• Shift from Prompting to Loops: Top AI engineers are moving away from manual prompting toward building loops that allow agents to work autonomously for hours.
• The Cost Barrier: Autonomous loops are extremely token-hungry; a single run for a simple app can consume over a million tokens.
• Deterministic Workflows: To ensure reliability, developers should use harnesses like Arkon to enforce specific processes rather than letting the LLM decide every step.
• Observability is Critical: Running long-term loops requires a dashboard or 'control plane' to monitor agent decisions, track costs, and allow for human-in-the-loop intervention.
• Model Mixing: Optimizing for cost involves using cheaper models (like Pi/Kimi) for orchestration and research while reserving expensive models (like Claude 3.5 Sonnet) for core implementation.

Detailed Summary

Introduction (0:00 - 01:42)

Loop engineering is becoming the next major buzzword in the AI space, championed by figures like Peter Steinberger and Boris Cherney. While the concept of agents working 24/7 is appealing, these systems are often unreliable and consume massive amounts of tokens. The goal is to move beyond the hype to build practical harnesses that provide observability and cost optimization.

The Core Concept of Loops (01:42 - 05:52)

In Claude Code, loop engineering is implemented through specific commands like /loop, /goal, and /routines.

• /loop: Sets an interval for a prompt to run (e.g., checking GitHub issues every 5 minutes).
• /goal: Forces an agent to work until specific criteria are met.
• /routines: Schedules jobs to handle larger specs incrementally.
• Orchestration: A primary agent acts as an orchestrator, using the 'loop skill' to spin up workers and manage task lists autonomously.

Downsides and Token Costs (05:52 - 08:34)

There are three primary problems with current loop engineering trends:

1. Practicality: Claims of managing tens of thousands of agents simultaneously are often hyperbolic and can lead to buggy output.
2. Extreme Expense: Without optimization, context passing between orchestrators and workers leads to massive token usage.
3. Context Bloat: Running loops in a single session overwhelms the LLM's context window; work must be distributed across separate sessions.

Deterministic Workflows with Arkon (08:34 - 13:18)

Arkon serves as a harness to solve reliability issues by making workflows deterministic. Instead of letting the agent drive the entire process, the developer defines a file-based workflow (e.g., Plan -> Implement -> Review). This allows for:

• Model Selection: Using smaller models for classification and larger ones for coding.
• Durability: Storing logs in a database (like Neon) so workflows can be resumed if a machine fails.
• Process Enforcement: Guaranteeing steps like testing occur regardless of agent 'reasoning.'

Orchestrating Parallel Coding Agents (13:18 - 17:29)

Scaling AI output requires running multiple agents in parallel using work trees to prevent them from interfering with each other's code or database changes.

• Parallelism: One orchestrator can dispatch multiple Arkon workflows to fix different GitHub issues simultaneously.
• Human-in-the-Loop: Systems should be designed to pause for human validation to prevent the 'loop' from generating large amounts of low-quality code.

The Pi Loop Engineering Dashboard (17:29 - 21:11)

A custom open-source 'Agent Control Plane' solves the state management problem by storing all loop events in an external Postgres database.

• State Management: The orchestrator reads state from the database, dispatches workers, and workers update the database, creating a clean feedback loop.
• Cost Efficiency: Using the Pi (Kimi) model allows for high-volume token usage at a lower price point while maintaining quality through the harness.

Deploying Control Systems to Production (21:11 - 23:58)

Local AI tools can be deployed to the cloud using platforms like Retool to enable team collaboration and remote access.

• Cloud Integration: Connecting the dashboard to a cloud database and API allows for governed paths to production.
• Permissions: Features like 'Approve and Resume' can be gated behind permission groups to ensure only authorized users can advance autonomous workflows.

Conclusion (23:58 - 24:37)

Loop engineering is essentially 'harness engineering.' While the buzzword may be temporary, the underlying principles of building durable, observable, and autonomous systems are the future of AI-assisted development.

Notable Quotes

"I don't prompt Claude anymore. I write loops and the loops do the work. My job is to write loops." — Boris Cherney (Head of Claude Code)

"We want to actually take the decision away from the coding agent as much as we can, only applying the reasoning of the LLM when we actually need it."

"Loop engineering is really not that complicated. I don't even know if it deserves its own term."