How to build an infinite agentic loop with two prompts

TL;DR

The video shows how to create an infinite agentic loop using only two prompts, enabling multiple sub agents to run in parallel. It demonstrates generating self-contained UIs and iterating through waves of parallel agents, while noting potential limits like context window capacity and how prompts can be passed into prompts. This pattern highlights scalable, self improving coding workflows.

Transcript

Engineers, I think I just broke Clawude Code, the best agent coding tool in the game. Check this out and you tell me. The amount of value you can create in a single prompt is incredible. The amount of value you can create in two prompts is insanely mindblowing and not well understood. Let me show you exactly what I mean. Here I'm running a clawed c... Read More

Key Insights

• The pattern uses two prompts to fuel an infinite agentic loop.
• Five sub agents run in parallel in the demonstration.
• Thousands of lines of code are generated across multiple waves.
• The loop uses an infinite prompt and a plan to guide work.
• Spec files are treated as first class citizens and passed to prompts.
• Context capacity is evaluated to decide whether to continue or finalize.
• Infinite agentic loops imply scalable but bounded computation.
• The approach shows how prompts can generate prompts for self improvement.

Questions & Answers

Q: How to set up an infinite agentic loop with two prompts?

The setup uses two prompts to drive parallel sub agents that act on a plan and a spec. The infinite loop is started by an initial command, which feeds the spec and iteration details to each sub agent. This structure enables continuous generation, with each wave producing new outputs and new prompts to maintain the loop while tracking iterations.

Q: What role do sub agents play in the loop?

Sub agents operate in parallel, each receiving a tailored prompt and iteration information. They work on their own tasks, producing output that contributes to the overall objective. The system organizes these agents in waves, monitors their progress, and queues additional agents to maintain momentum and scale the workflow without manual intervention.

Q: Why are specs treated as first class citizens in this pattern?

Treating specs as first class citizens allows prompts to be dynamic and self modifying. By passing the spec file into prompts, agents can adapt their behavior based on evolving requirements. This enables a more flexible, rule driven workflow where outputs can influence subsequent prompts and iterations in a structured way.

Q: How does context window capacity affect the loop?

Context window capacity acts as a practical constraint. The pattern includes a check to evaluate available context and decide whether to continue with the next wave or finalize the work. Once the context limit is approached, the system terminates or consolidates results to prevent overflow and maintain coherence of the generated content.

Q: What is the purpose of the infinite prompt in this demo?

The infinite prompt is the core driver that enables continuous, self orchestrated generation. It defines how prompts should be combined, how outputs should be organized, and how to spawn parallel agents. This structure allows iterative improvement and expansion across multiple iterations while maintaining alignment with the specified goals and plans.

Q: What is meant by the term agentic loop in this context?

An agentic loop refers to a repeating cycle where agents generate outputs, which in turn guide the creation of new agents and new prompts. This loop creates a self accelerating pattern where solutions evolve through successive waves of parallel computation, enabling the exploration of many possibilities concurrently.

Q: What are potential risks of running an infinite agentic loop?

Potential risks include overwhelming resource usage, loss of coherence as outputs proliferate, and exceeding context limits. There is also risk of degraded safety if prompts begin to generate unsafe or undesirable content. Proper safeguards and termination criteria help prevent runaway execution and ensure outputs remain controllable and useful.

Q: What practical applications are suggested by the video for this pattern?

The video suggests real world use cases where infinite solution generation and parallel exploration could accelerate engineering workflows. By combining planning with principled prompting, teams could rapidly generate multiple design concepts, UI patterns, or code scaffolds, iterating on improvements as prompts produce new variations and refinements in parallel.

Summary & Key Takeaways

• The video presents an infinite agentic loop pattern driven by two prompts, enabling rapid, parallel generation of self improving outputs. It emphasizes treating prompts as first class citizens and passing specs into prompts to coordinate multiple agents. It also discusses practical constraints like context window limits and terminating when nearing limits.

• A practical demonstration shows five sub agents in parallel writing thousands of lines of code, producing a stream of iterations and new UI components. The author explains the setup, prompts, and output directories, illustrating how parallelism accelerates creative coding tasks within Claude Code.

• The core takeaway is a scalable workflow where prompts generate more prompts, enabling an ongoing cycle of solution generation and refinement. The video also cautions about limitations and the need to manage resources and context when running infinite agentic loops.