SolidState is an open-source runtime for building safe and controllable AI agents. It separates probabilistic reasoning from deterministic execution, giving the runtime control over state transitions, tools, approvals, retries, concurrency, persistence, and limits. The alpha release includes FSM workflows, human-in-the-loop policies, parallel tools, checkpointing, StateGraph orchestration, actor-based concurrency, tracing, and multi-model support.

I built SolidState because I kept seeing the same problem in agentic AI systems: the model was being asked to reason and control execution at the same time. LLMs are excellent at interpreting intent, planning steps, and proposing actions, but production systems need deterministic behaviour around tool access, retries, approvals, state transitions, concurrency, and failure handling. Prompt-level guardrails alone are not enough when an agent can modify data, call external services, or trigger business operations. That led to the core principle behind SolidState: **The model reasons. The runtime controls execution.** SolidState uses a deterministic finite state machine to manage the execution lifecycle around the model. It provides typed runtime state, bounded loops, policy-based Allow/Reject/Interrupt decisions, human approval workflows, checkpointing, parallel tool execution, multi-agent StateGraph orchestration, tracing, and actor-based concurrency. The design evolved from a set of architecture discussions and implementation experiments into a working open-source framework. The current alpha release is intentionally focused on making runtime behaviour explicit, inspectable, and extensible rather than hiding everything behind another abstraction layer. This is still an early release, so feedback is especially valuable. I would love to hear from developers, architects, and researchers working on agent runtimes, AI governance, multi-agent systems, and production LLM applications. 🌐 Website: [https://solidstatehq.ai](https:/... 💻 GitHub: [https://github.com/gavarah/solid... 📦 PyPI: [https://pypi.org/project/solidst... 📄 Technical report: [https://zenodo.org/records/21401... Install with: pip install solidstate-fsm

The split between probabilistic reasoning and deterministic execution actually clicked for me once I set up a workflow with an approval gate. Checkpointing saved me when a flaky tool call blew up midway.

Tried the alpha and the FSM workflow setup is genuinely intuitive for something open source. Really like that approvals and retries are baked in rather than bolted on.

SolidState HQ

A deterministic runtime for safe, controllable AI agents

A deterministic runtime for safe, controllable AI agents