Stop Using AI Agents Like Chatbots. (Do This Instead)

Building Software with Coding Agents: A Practical Workflow

Key Concepts:

• Coding Agents: AI systems designed to assist or automate software development tasks.
• Verdant: A specific coding agent featuring parallel work via Git trees, a clean UI, and Plan/Agent modes.
• Plan-Execute-Test-Verify Workflow: A cyclical process for building software incrementally with coding agents.
• MCP (Multi-Code Plugins): Tools integrated into coding agents to extend their capabilities (e.g., ChromeDevtool for web app verification).
• Git Trees/Workspaces: A method for managing parallel development branches without affecting the main codebase.
• Context Window: The amount of text a language model can process at once (Verdant appears to compact conversations, potentially limiting effective context).
• Sub-agents: Specialized agents used for specific tasks within a larger workflow (e.g., testing, verification).
• Slash Commands: Custom commands that can be used within a coding agent to trigger specific actions or skills.

I. The Shift in Software Building & The Importance of the Harness

The video highlights a significant shift in software development: the harness around a coding agent – the workflow, tools, and environment – is becoming more crucial than the raw intelligence of the underlying model. Regardless of the specific coding agent used, a consistent workflow yields better results. The speaker emphasizes that the input method (dictation software) is currently a bottleneck in the process.

II. Introducing Verdant & Its Key Features

The demonstration focuses on Verdant, a coding agent the speaker participated in the early access program for (and which is sponsoring the video). Verdant distinguishes itself through:

• Parallel Work: Utilizing Git trees and workspaces, Verdant allows for simultaneous implementation of multiple features without code conflicts. New workspaces create isolated branches using git workspace.
• Clean UI: A streamlined interface designed to minimize distractions and focus on ideation.
• Plan & Agent Modes: A dual-mode system. Plan Mode is for initial brainstorming and refining ideas (powered by Opus 45 in the example), while Agent Mode handles code implementation.
• Skill Integration (Workaround): While lacking direct skill integration like Cloud Code, Verdant supports custom commands to achieve similar functionality.

III. The Plan-Execute-Test-Verify Workflow: A Practical Example

The core of the video is a demonstration of the Plan-Execute-Test-Verify workflow, building a text-to-image generator using Google’s Nano Banana model.

Step 1: Plan (Using Verdant’s Plan Mode)

• Initial Prompt: "Help me create a text to image generator. We're going to use Google Nano Banana model to generate four different images."
• Refinement: The prompt was refined to request HTML, CSS, and JavaScript. The speaker requested a design aesthetic "designed by a billion-dollar design company."
• Detailed Plan Generation: Verdant generated a detailed plan outlining API endpoint usage (specifically the REST API, avoiding SDKs), image generation, and display logic.
• Definition of Done: The plan included a clear definition of completion criteria.
• Model Used: Opus 45 was used for both planning and implementation.

Step 2: Execute (Switching to Verdant’s Agent Mode)

• Automatic Mode Switch: Clicking "Build" automatically transitioned from Plan Mode to Agent Mode.
• Parallel Task Creation: The speaker demonstrated creating a second task in parallel, illustrating Verdant’s workspace feature.
• Code Implementation: Verdant autonomously generated the code.

Step 3: Test & Verify

• Initial Testing & Rate Limits: The initial test generated only one image, likely due to API rate limits.
• Automated Verification with MCP: The speaker integrated the ChromeDevtool MCP server to enable automated web app verification. This involved:
  • Accessing settings and enabling the ChromeDevtool.
  • Verdant autonomously writing a prompt and executing it within the Chrome Devtool.
  • Identifying and automatically fixing code issues.
• MCP Tool List: Verdant provides a list of available tools within the MCP server.

Step 4: Incremental Feature Addition & Design Improvement

• Custom Skill Implementation (Workaround): To address the lack of direct skill integration, the speaker created a custom command ("frontend design") by copying the markdown text from a Cloud Code design skill.
• Parallel Feature Development: A second task was created to implement a dark/light theme toggle, running concurrently with the design improvement task.
• Automated Verification of New Features: The ChromeDevtool was again used to verify the functionality of the dark/light theme toggle.

IV. Addressing Limitations & Best Practices

• Context Window Compaction: The speaker noted that Verdant appears to compact conversations, potentially limiting the effective context window (200,000 tokens). This reinforces the need for incremental feature development.
• Avoiding Simultaneous Feature Implementation: Implementing too many features at once increases the "surface area for failure."
• Minimal Sub-agents: The speaker prefers to limit the use of sub-agents, primarily utilizing them for verification and testing.
• Dictation Software: The speaker recommends using dictation software to overcome the input bottleneck. (Link to the speaker’s dictation system is in the video description).

V. Notable Quotes

• “The harness around your coding agent is probably more important than the raw intelligence of the model itself.” – Emphasizing the importance of workflow and tooling.
• “I usually like to use these coding agents as brainstorming partners.” – Highlighting the value of agents in the ideation phase.
• “In this case we had to manually verify the outputs but it's 2026. We shouldn't be doing that.” – Expressing the expectation of increased automation in verification.

Conclusion:

The video advocates for a simple yet effective workflow – Plan, Execute, Test, Verify – when building software with coding agents. The demonstration with Verdant highlights the benefits of parallel development, automated testing, and incremental feature addition. The key takeaway is that a well-defined process, combined with the right tools (and a workaround for missing features like skill integration), can significantly enhance the efficiency and quality of software development using AI-powered agents. The speaker stresses the importance of focusing on the harness – the workflow and environment – rather than solely relying on the raw power of the underlying model.