zenflow: Design and Implementation of a Declarative Multi-Agent Orchestration Engine

zenflow is a production-grade multi-agent workflow engine based on Go. It achieves multi-agent collaborative orchestration capabilities through declarative YAML configuration, a central coordinator architecture, and a reliable mailbox delivery mechanism. It aims to balance simplicity and functionality, addressing engineering challenges in multi-agent systems such as workflow definition, reliable message delivery, and concurrent processing.

With the improvement of LLM capabilities, multi-agent collaboration systems face many challenges: How to define complex workflows? How to ensure reliable message delivery? How to handle concurrent race conditions? How to avoid data loss in case of failures? Existing solutions are either too simple (only linear execution) or too complex (relying on heavyweight infrastructure), and zenflow attempts to fill this gap.

zenflow adopts a hub-and-spoke architecture. All agent communications are forwarded through a central coordinator, avoiding direct coupling and simplifying dependency management and fault isolation. The coordinator is responsible for message routing, progress tracking, event forwarding, and execution finalization; each agent has an independent mailbox to enable explicit delivery confirmation and failure tracking.

zenflow uses YAML to define workflows, supporting primitives such as steps, dependencies, parallel fan-out, CEL conditional branches, loops, and sub-workflows. The declarative design facilitates version control and collaboration. Strict validation (cycle detection, dependency integrity, CEL expression validation) is performed before execution to ensure issues are identified early.

zenflow ensures no message loss, no out-of-order messages, and no goroutine leaks through race-safe mailbox delivery. It has a built-in failure recovery mechanism that resumes execution based on persisted state, ensuring completed steps are not repeated and incomplete steps continue from breakpoints, achieving zero data loss.

zenflow provides three execution modes:

1. Flow mode: Runs deterministic YAML DAGs, suitable for scenarios with clear plans;
2. Goal mode: Dynamically plans workflows based on user input, adapting to scenarios requiring adaptive adjustments;
3. Agent mode: Single-agent dialogue, reusing lifecycle hooks and provider routing. It also supports embedding integration as a Go library.

zenflow is compatible with mainstream LLM providers such as Google Gemini, AWS Bedrock, and Azure (all providers supported by the goai library can be used); it offers a sandbox mode to restrict dangerous tools (e.g., bash), suitable for automation scenarios. Deployment is released as a single static Go binary, supporting multiple methods such as one-click installation, Homebrew, Docker, and Go install.

zenflow represents the trend of multi-agent orchestration tools evolving toward production-grade. It lowers the threshold through declarative configuration, simplifies design with a central architecture, and improves reliability via strict validation. For developers building complex agent workflows, it is a worthy engineering solution to consider.