OpenClaw
Decision models give OpenClaw plugins a way to make fast, typed choices without routing every judgement through a conversational model. Developers can configure a separate model, then call it from application code through the Plugin SDK.
The OpenClaw blog post explains the problem this is meant to solve. Agents make many small decisions while handling a request. They may need to choose relevant tools, decide which messages survive compaction or work out whether somebody is speaking to the agent.
A language model can make those calls, but each request adds time and cost to the user's main task. It can also return a paragraph when the application needs a specific value.
How Decision Models Work in OpenClaw
A decision model takes evidence and a set of criteria. It returns a typed answer that application code can use immediately, such as a choice, score or probability.
| Approach | How it is called | What it returns |
|---|---|---|
| Conversational model | The agent decides when to call it as a tool | A language response that may need interpretation |
| Decision model | Application code calls it directly | A typed choice, score or probability |
| OpenClaw plugin | Calls the configured decision model through the Plugin SDK | An evaluation for the plugin's next step |
Plugins access the configured model through api.runtime.decisions.evaluate. This keeps the decision inside normal deterministic code rather than asking the conversational model to decide whether it should call another model.
OpenClaw is taking a plugin-first approach. Its initial TypeSafe adapter works with Jev, the decision model created by Diogo Almeida and the TypeSafe team. The adapter supports both hosted Jev and a local System One server.
Why OpenClaw Is Using Plugins
Josh Lehman, an OpenClaw maintainer, first tested Jev by giving his agent a tool that could call it. That worked, but the slower language model still had to decide when to use the faster API. Calling the decision model directly from application code removes that extra step.
The shared SDK method also gives other plugins one interface for requesting decisions. OpenClaw is inviting contributors to build on the capability through its plugin system.
Frequently asked questions
What are Decision models in OpenClaw?
They are models that evaluate evidence against defined criteria and return a typed choice, score or probability. OpenClaw plugins can call the configured model directly through the Plugin SDK.
Is OpenClaw Decision models free?
The post does not give pricing or plan details for Decision models. It says the TypeSafe adapter supports hosted Jev and a local System One server.
How do Decision models compare with conversational models?
Conversational models produce language responses and add another model call to the agent's task. Decision models return a focused, typed result that application code can use for its next step.
Sources
3 checkedHow we cover tool news: Create With's tool desk drafts these reports with AI from the sources listed above and checks them against those sources before publishing.
Momen
Vercel
Make