AgentCastle: A Kanban-based Pi Agent System Integrating Token-Efficient Tools and Secure Sandboxes

AgentCastle is an open-source project developed by SchneiderDaniel, with its core being a Kanban-centric Pi Agent system. It integrates token-efficient tool design, multi-layered security guardrails, and Docker sandbox execution environments to enable autonomous Kanban pipeline management, providing a safe and reliable execution framework for AI Agent workflows. Project URL: https://github.com/SchneiderDaniel/agentcastle, Release Date: 2026-05-25.

In today's era of rapid AI Agent development, building both efficient and secure Agent systems has become a key challenge. The AgentCastle project deeply integrates project management methodology (Kanban) with AI Agent technology to propose an innovative solution. Its core features include token-efficient tools, multi-layered security protection, Docker sandboxes, and real-time feedback mechanisms, aiming to address the efficiency and security issues of Agent systems.

Pi Agent Architecture

"Pi" stands for Pipeline Intelligence, Process Intelligence, and Portable Intelligence. Its core is to encapsulate AI capabilities into composable modular components, managed via Kanban.

Kanban-Centric Design

• Visual workflow: Intuitively display task status
• WIP (Work in Progress) limits: Prevent system overload
• Flow optimization: Eliminate bottlenecks
• Pull system: Agents actively pull tasks

Technical Architecture

• Token-efficient tools: Concise descriptions, structured outputs, intelligent context management, tool combination optimization
• Multi-layered security protection: Input validation, permission control, behavior monitoring, audit logs
• Docker sandbox: Complete isolation, resource limits, consistent environment, fast recovery
• Real-time feedback: Status streams, log push, error alerts, performance monitoring

Workflow

1. Task Creation and Board Entry: Natural language description, template selection, batch import; tasks enter the "To Do" column
2. Agent Autonomous Pull: Capability matching, load awareness, priority sorting; tasks move to "In Progress"
3. Sandbox Execution: Environment preparation, tool invocation, security monitoring, result collection
4. Result Verification and Completion: Automatic verification/manual review; tasks move to "Completed" column, knowledge is accumulated

Application Scenarios

• Software development: Code review, test execution, document generation, dependency updates
• Data engineering: Pipeline monitoring, quality inspection, ETL tasks, report generation
• Operations: Alert response, log analysis, backup verification, configuration management
• Content creation: Content review, SEO optimization, multi-platform publishing, performance tracking

Technical Advantages

• Kanban-AI integration: Provides human-machine collaboration interface, high transparency, strong flexibility
• Security-first: Zero-trust architecture, least privilege principle, defense in depth
• Observability: Multi-dimensional metrics, log aggregation, end-to-end tracing

Comparison with Similar Projects

• Native Kanban: Core design rather than a plugin
• Pi Agent concept: Emphasizes process and pipeline intelligence
• Built-in Docker sandbox: No need for external security mechanisms
• Token optimization: Efficiency considered at the architectural level

Future expansion directions for AgentCastle:

1. Multi-Agent collaboration: Support multi-Agent coordination for complex tasks
2. Learning optimization: Learn from historical executions to improve decision-making
3. Integration expansion: Provide more out-of-the-box tools
4. Mobile support: Develop mobile management applications
5. Community ecosystem: Establish a plugin market and encourage community contributions