Agent Conductor: Transforming Agent Capabilities into Verifiable, Visualizable Workflow Orchestration System

Agent Conductor is a set of lightweight specifications and a local kanban tool designed to address the trust crisis in LLM Agent workflows. It transforms any agent capability into a card-based workflow with independent verification steps, enhancing the reliability and observability of AI task execution through structured instructions and dependency management. Its core goal is to turn complex agent tasks from black-box execution into white-box orchestration, enabling controllable and auditable workflow management.

As LLM-driven agents move from experimentation to production, the tasks they perform are becoming increasingly complex (e.g., research, report writing, code review). However, traditional software testing methods are difficult to apply to the open-ended natural language outputs of agents. For example, an agent might fail to complete the required data collection or cite unreliable information, and users have to read the entire output to detect problems. This trust crisis led to the design of Agent Conductor.

The core concepts of Agent Conductor include:

1. Card-based Decomposition: Split tasks into clear card units, each containing an ID, title, instructions, and dependencies, forcing ambiguous goals into verifiable steps.
2. Independent Verification: After each card is completed, it must pass independent verification (manual review, automated scripts, or external tools). Verifiers are separated from executors to avoid self-rating bias.
3. Visual Tracking: The local kanban displays card statuses (Pending/Running/Checking/Done/Failed) in real time, enhancing task observability.

Agent Conductor defines workflows via conductor.json, which includes top-level fields (version, name, description, maximum retries, step array) and card fields (ID, title, instructions, dependencies). The dependency graph supports complex execution sequences. The CLI tool conductor-board provides commands:

• check: Inspect card outputs
• complete: Mark a card as completed (verification required first)
• gate-result: Record verification results and evidence
• Commands for initializing projects, validating workflows, etc.

Agent Conductor is suitable for various scenarios:

1. Content Creation: An ordered workflow of research → outline → draft → editing → fact-checking.
2. Code Review: A combination of parallel and serial steps: static analysis → security scanning → test coverage → manual review.
3. Data Pipeline: ETL process of extraction → transformation → verification → loading, with verification checkpoints at each stage.

Design Philosophy:

• Explicit over implicit: Clearly define step inputs and outputs, dependencies, and verification methods.
• Separation of verification and execution: Avoid self-rating bias.
• Human-in-the-loop: Support human intervention for adjustments.
• Tool-agnostic: Compatible with any agent framework. Limitations:
• Learning curve: Requires mastering the specifications and task decomposition.
• Verification cost: Additional overhead from manual or automated verification.
• State management: State file synchronization is needed for multi-person collaboration.

Agent Conductor turns agent workflows from black-box to white-box through card-based decomposition, independent verification, and visual tracking, enhancing reliability and auditability. It does not provide more powerful agents but makes the use of existing agents more reliable. Teams that are putting agents into production are recommended to try this system; its learning curve is moderate, and the improvements in observability and verification capabilities are significant.