Skip to main content

Agent-Agnostic Task Orchestration System

Kiro-Style Workflow Without Vendor Lock-in​

This system implements a robust, standardized interface for task management across multiple AI agents. By treating agents like interchangeable backends, we eliminate vendor lock-in and ensure JACK's development can scale regardless of which tool is currently leading the market.

Core Philosophy: The Agent Interface Pattern​

The system follows a tiered architecture that separates planning from execution:

Key insight: Treat agents like HTTP backends - standardize the interface, swap implementations.

The Kiro Pattern​

We adopt the highly effective Kiro workflow:

  1. Requirement document: Clear definition of what and why.
  2. Task breakdown: Atomic, verifiable units of work.
  3. Agent execution: Autonomous or assisted processing.
  4. Verification: Automated checks (compilation, tests, linting).
  5. Feedback loop: Retries or adjustments based on failure.

How to Use​

1. Define Your Tasks​

Create a YAML file in .agent-tasks/ following the schema.

2. Run the Orchestrator​

# Execute all tasks in a file
pnpm agent:run .agent-tasks/day-1.yaml

# Target a specific agent
PREFERRED_AGENT=kiro pnpm agent:run .agent-tasks/day-1.yaml

3. Verification & Commitment​

The system automatically:

  1. Runs verification scripts (e.g., forge build).
  2. Creates a git commit with the agent's log.
  3. References and closes relevant GitHub issues.

Monitoring (The Agent Dashboard)​

Use the TUI dashboard to monitor multiple agents working in parallel:

pnpm agent:dashboard

Cost & Performance Tracking​

Every execution logs:

  • Duration: How long the agent took to solve the task.
  • Success Rate: Automated verification results.
  • Cost: Estimated USD/Token consumption per agent.

This data allows the team to optimize agent selection for specialized tasks (e.g., using cheaper agents for boilerplate and premium ones for security logic).