A Security Paradigm for Agentic Autonomy.
SoyAgent is designed to solve the fundamental challenge of modern AI architecture: granting autonomous systems deep tool access without expanding the attack surface.
Isolation & Credential Layer
Secret Injection Proxy
The reasoning engine generates unauthenticated intent. The proxy intercepts egress, terminates TLS, and injects Authorization headers from an isolated Vault.
Linux Mount Namespaces
Physical isolation at the kernel level. Bind-mount only explicitly approved directories. Sensitive host files remain invisible to the agent.
AgentFS (SQLite Virtualization)
All file operations are routed to SQLite. Modifications create differential deltas, ensuring host files are never overwritten directly.
Orchestration & Persistence
Plan-and-Execute DAGs
State SerializationSeparate cognition from execution. A Director Agent builds a structured plan, which is compiled into a deterministic LangGraph state machine.
Durable Background Workers
Trigger.dev IntegrationTasks checkpoint state after every step. If the server crashes, SoyAgent resumes precisely where it left off without duplicating work.
Decoupled Authorization (CIBA)
SoyAgent uses the **Client-Initiated Backchannel Authentication** standard to manage Human-on-the-Loop approvals. When a sensitive tool is invoked, the agent doesn't block—it hibernates.