- Plan - Before anything is executed, talk through a clarification loop with the AI. You can customize its behavior with plan rules, and use interactive architecture planning, visual diagrams, and task breakdowns to make your requirements crystal clear
- Code - Review changes with Code Diff’s enterprise-grade analysis and deliver production-ready results with confidence
- Verify - Quality assurance with Verify subagent and browser action. Iterate through multipass repair cycles until code passes comprehensive validation
Architecture Overview
Complex tasks are broken down and delegated to specialized AI agents optimized for specific purposes. Each subagent operates with its own custom system prompt and separate context window, isolated from the main agent that invoked it. This separation maintains peak performance and prevents context pollution, allowing each subagent to focus on its specialized task without burdening the main agent with every implementation detail. Built-in Default Subagents: Verdent includes three default subagents immediately available:| Subagent | Purpose | When It’s Used |
|---|---|---|
| @Verifier | Quick code checks and validation | During coding tasks to ensure code quality and correctness |
| @Explorer | Fast codebase exploration | Finding files and answering questions about your codebase structure |
| @Code-reviewer | Instant code quality assessment | Proactive scanning for security, maintainability, and performance issues |
- Technology Stack Specialists - React performance optimizer, SQL query optimizer, Docker configuration reviewer
- Quality Assurance - Test coverage analyzer, error handling reviewer, logging standards enforcer
- Team-Specific Workflows - Code style enforcer, documentation consistency checker, dependency auditor
- Project-Specific Needs - Legacy code modernizer, migration assistant, performance budget enforcer
Execution Modes
Verdent supports multiple execution modes that control how the AI handles permissions and interactions: Permission Modes:- Manual Accept Mode - Requests permission on first use of each tool. Best for learning and general development with maximum control.
- Auto-Run Mode (default)- Automatically accepts permissions for safe operations (file reads, web searches). File edits and command execution still require approval. Ideal for trusted projects and rapid prototyping.
- Skip Permissions Mode - Fully autonomous operation bypassing all prompts. Only for isolated, safe environments like automated testing or sandboxes.
- Agent Mode - Full agentic behavior with autonomous task execution and file operations
- Plan Mode - Planning-focused interaction for reviewing approaches before execution
- Chat Mode - Conversational interaction without autonomous file operations
- Think Hard Mode - Allocates maximum computational resources to analyze problems deeply, explore multiple approaches, and provide more thorough solutions. Ideal for architectural decisions, complex debugging, and tasks requiring sophisticated reasoning.
Context Management
Context management controls what information Verdent has access to during your session. The context window is the agent’s memory limit, measured in tokens and shared across several components: system instructions, file contents, conversation messages, and agent responses. How Context Works:- @-Mentions - Explicitly add files to context using
@filename.jsfor precise control - Automatic Loading - Verdent automatically includes relevant files based on the task
- Separate Subagent Context - Each subagent operates with isolated context windows, preventing pollution of the main agent’s memory
- Context Window Limits - Finite context measured in tokens. As conversations grow, available space decreases and model performance degrades
- Break large tasks into smaller chunks with natural breakpoints
- Use Plan Mode for complex tasks to review approaches before execution
- Complete individual components before integration to avoid context depletion
- Start a new conversation when beginning a new task to clear context and maintain optimal performance
Model Selection
Verdent orchestrates models from multiple AI providers (Claude, GPT-5, Gemini, MiniMax) within a single workflow. Each specialized agent can leverage the optimal model for its specific task. Preset Models:- Performance - Maximum context, frontier models (1-2x credit usage)
- Balance - Optimal blend of performance and cost (1x credit usage)
- Efficiency - Fast, responsive, cost-efficient tasks (~0.3x credit usage)
| Provider | Model | Best For | Cost Efficiency |
|---|---|---|---|
| Anthropic | Claude 4.5 Opus | Great for deep coding tasks and problem-solving. | 1.5x cost |
| Anthropic | Claude 4.5 Sonnet | Balanced complex tasks | 1x baseline |
| Anthropic | Claude 4.5 Sonnet 1M | Long or complex tasks (>200k input) | 0.5x efficiency (2x cost) |
| Anthropic | Claude 4.5 Haiku | Fast, simple tasks | 3.2x more efficient than Sonnet |
| OpenAI | GPT-5 (Beta) | Reasoning and creativity | 1.3x more efficient than Sonnet |
| OpenAI | GPT-5 Codex (Beta) | Coding and debugging | 1.3x more efficient than Sonnet |
| Gemini 3 Pro | Advanced reasoning and multimodal | Variable | |
| Kimi | Kimi 2 Turbo (Beta) | Lighting-fact model for coding | 3.1x more efficient than Sonnet |
FAQs
What is context pollution?
What is context pollution?
When previous actions and information in the context create unintended behavioral patterns that affect future tasks. For example, if you repeatedly update code and then deploy it, the agent may begin associating all code updates with immediate deployment, even when you’re just experimenting. Subagents prevent context pollution by operating with isolated context windows, allowing each specialized task to start with clean context and avoiding cross-contamination between different types of work.
What are safe operations?
What are safe operations?
Operations that only read data without making changes: file reads and web searches. These are automatically approved in Auto-Run Mode, while file edits and command execution still require approval for safety.
What is command execution?
What is command execution?
Running terminal commands or shell scripts through Verdent, including operations like running tests, installing packages, or executing build commands. Command execution requires approval in most Permission Modes to prevent unintended system changes.
What is a token?
What is a token?
The unit of measurement for context windows. Tokens represent pieces of text (words, parts of words, or characters) that the AI model processes. Context limits are measured in tokens rather than characters or words, with typical conversations consuming thousands of tokens.
What are credits?
What are credits?
The usage currency for Verdent. Different models have different cost efficiency (e.g., Haiku is 3.2x more efficient than Sonnet baseline, Sonnet is 1x baseline). Your subscription includes a monthly credit allotment (e.g., 640 credits for Starter, 2,000 for Pro, 6,000 for Max).
What is the Multipass Generate-Test-Repair Cycle?
What is the Multipass Generate-Test-Repair Cycle?
An iterative verification process where code is generated, tested, and repaired until it passes comprehensive testing. Verdent automatically runs tests, analyzes failures, fixes issues, and retests in multiple passes until the code works correctly. This delivers production-ready code, not just suggestions. Each result includes detailed summaries and code diffs showing exactly what changed.