Building the App Store for Agentic Engineering

Key Concepts

• Archon: An open-source "harness builder" for AI coding that allows users to package agentic engineering processes into reliable, repeatable, and parallelizable workflows.
• PIV Loop: A core development methodology (Plan, Implement, Validate) used by the creator to ensure alignment between human intent and AI execution.
• Workflow Marketplace: A new feature for Archon that allows users to share custom workflows via GitHub pull requests, which are then automatically reviewed and merged.
• Second Brain: The creator’s primary development environment (a repository managed by Claude Code) used to orchestrate Archon workflows.
• Work Trees: A Git feature used by Archon to isolate codebases, allowing multiple workflows to run in parallel without interference.
• Agentic Coding: The process of using AI agents to perform complex software development tasks, requiring careful context management and human-in-the-loop oversight.

---

1. Main Topics and Key Points

The live stream focused on building a Workflow Marketplace for Archon. The goal is to move away from a "black box" tool toward a platform where users can share and customize their own AI coding processes.

• Self-Evolving System: The marketplace uses an Archon workflow to automatically review and approve pull requests (PRs) that add new workflows. This reduces the maintenance burden on the project creator.
• Automation Strategy: The system uses a pull_request_target GitHub Action trigger. This allows the workflow to run in the context of the base repository (with access to necessary secrets like the Anthropic API key) while safely reviewing PRs from forks without executing untrusted code directly on the runner.
• Token Efficiency: The creator emphasizes using different models for different tasks (e.g., Claude Haiku for cost-effective reviews, Sonnet/Opus for complex implementation) to manage API costs.

2. Step-by-Step Process: The PIV Loop

The creator utilizes a structured framework for building features:

1. Exploration (Prime Phase): The agent scouts the codebase and loads relevant context.
2. Planning: The agent generates a structured Markdown plan. A human-in-the-loop approval gate is mandatory here to prevent "one bad line in a plan leading to a thousand bad lines of code."
3. Implementation: The agent executes the plan in an isolated Git work tree.
4. Validation: A separate, fresh coding agent session performs a code review, comparing the implementation against the original plan and running tests.

3. Key Arguments and Perspectives

• Harness Builder vs. Harness: The creator argues that Archon is superior to rigid "harnesses" (like BMAD) because it is not opinionated. It allows users to define their own processes rather than forcing them into a pre-built structure.
• Reliability through Determinism: The creator posits that "prompting is brittle." By building Archon workflows, users move from "system-in-the-agent" (where the agent must remember all instructions) to "agents-in-the-system" (where the system enforces the process step-by-step).
• Security: While the marketplace is community-driven, the creator acknowledges the risk of supply chain vulnerabilities. The current mitigation is a combination of automated AI reviews, a "community-submitted" disclaimer, and the recommendation to only run workflows from trusted authors.

4. Notable Quotes

• "When you build a skill for your AI coding workflow, you are building the system into the coding agent. When you are building an Archon workflow, you're building coding agents into a system."
• "I don't want to just build it from scratch. It's really useful to actually have Archon help me create the YAML so that I don't always have to remember the exact syntax."

5. Technical Details and Observations

• Tooling: The creator uses Aqua Voice for speech-to-text, noting that he can "brain dump" at 224 words per minute, significantly faster than his typing speed (106–120 WPM).
• Debugging: The stream highlights the reality of "agentic coding," where the AI occasionally hallucinates commands (e.g., using archon chat instead of archon workflow approve). The creator treats these as bugs to be fixed in the Archon codebase rather than failures of the concept.
• Orchestration: The "Second Brain" repository acts as the primary orchestrator, dispatching tasks to the Archon CLI, which then handles the heavy lifting in isolated environments.

6. Synthesis and Conclusion

The live stream successfully demonstrated the implementation of an automated marketplace for Archon. By the end of the session, the infrastructure for submitting, reviewing, and merging workflows was established. The creator plans to refine the security scanning and finalize the marketplace UI (archon.diy/workflows) for the next release. The session served as a practical, "in-the-trenches" look at how to build complex, agent-driven systems while maintaining human control and process reliability.