AxonFlow: Preventing Side Effects from Agent Retries

When an AI agent uses tools to modify a file system or call external systems, if an error occurs mid-workflow and a retry is performed, this can cause side effects where already-completed tasks are executed again.

AxonFlow is a source-available (BSL 1.1) runtime designed to solve this problem.

• State Recording and Resumption: It stores completion records for each step, allowing retries to resume from the point of failure instead of replaying the entire workflow.
• Local-First Design: It operates without cloud dependencies and is compatible with various local inference engines such as Ollama, vLLM, and llama.cpp.
• Inline Policy Checks: It checks for PII (personally identifiable information), SQL injection, rate limits, and more before each model call, recording a low latency of 7ms (P95) in tests.
• Tech Stack: It consists of two Go binaries, making it lightweight and independently executable.