CodeSoul-co/Hypha

What it solves

Hypha provides a production-ready framework for building governed, durable, and reusable domain-specific AI agents. It solves the problem of moving agents from simple prototypes to production by separating the core reasoning logic from the execution harness, ensuring that agent actions are bounded, auditable, and recoverable.

How it works

Hypha uses a two-layer architecture consisting of an Agent Core for reasoning, planning, and tool selection, and a Production Harness that manages the actual execution via a Finite State Machine (FSM) with built-in policy and approval gates.

Product-specific logic is not hard-coded but is instead defined in a DomainPack—a configuration that compiles tasks, workflows, prompts, and capabilities into the shared runtime. To optimize performance and cost, a Cache & Reuse Plane manages multiple layers of cached data (such as reasoning subgraphs and tool results) without allowing the cache to override the source-of-truth event logs.

Who it’s for

It is designed for developers building specialized agents for industries like finance, legal, or research, where reliability, auditability, and strict governance of tool execution are required.

Highlights

• DomainPack System: Decouples product-specific behavior from the runtime, allowing one framework to support multiple different agent types.
• Event-First Runtime: Uses an event-backed execution model to enable recovery, replay, and full auditing of agent actions.
• Multi-Layered Caching: Implements specialized cache trees (e.g., PlanTree, ToolTree) to reuse validated work and reduce LLM costs and latency.
• Governed Capabilities: Integrates tool and MCP (Model Context Protocol) adapters through a governed invocation path with policy and approval controls.
• Durable Execution: Supports checkpoints, bounded continuation, and recovery workers to ensure long-running tasks are not lost.