Overview
Stream of Thought
Step-by-step display of AI reasoning process
reasoningtransparencysteps
Props
| Prop | Type | Default | Description |
|---|---|---|---|
steps | ThoughtStep[] | required | Array of thought steps with id, type ("thinking" | "action" | "tool_call" | "result"), content, timestamp, and optional metadata |
isStreaming | boolean | false | Whether new steps are still being streamed in |
collapsible | boolean | false | Allow the thought stream to be collapsed/expanded |
Docs
Stream of Thought makes the AI's internal reasoning visible to the user. Instead of presenting only the final answer, it reveals the plan formed, tools invoked, code executed, and decisions made along the way. This transparency transforms a black-box interaction into an auditable process where users can verify logic, catch mistakes early, and build trust in the system's capabilities. The level of detail can scale from a brief progress summary for simple tasks to a full execution trace for complex, multi-step workflows.
Variants
- Human-readable plans — A preview of the steps the AI intends to take, shown before execution begins. Users can review, edit, or approve the plan.
- Execution logs — A real-time record of tool calls, code runs, and intermediate results as the AI works through a task.
- Compact summaries — Condensed accounts of reasoning, key insights, and decisions — useful when full logs would overwhelm the user.
Use Cases
- Multi-step agentic workflows where the AI calls tools or APIs
- Debugging and code generation tasks where reasoning matters
- Research and analysis where source evaluation is important
- Compliance and audit scenarios requiring decision traceability
- Educational contexts where users want to learn from the AI's process
Best Practices
- Show the plan before acting — let users review intended steps, estimated scope, and required permissions before execution begins.
- Keep three layers distinct: what will happen (plan), what is happening (execution), and what supports the result (evidence). Sync them but don't merge them.
- Scale detail to task complexity: a simple question needs minimal trace, while a multi-step workflow benefits from a full log.
- Treat each step as a clear state — queued, running, waiting for approval, error, retried, or completed — and pair states with visual progress cues.
- Adapt the presentation to the medium: text interfaces can link outputs to source steps, voice interfaces should summarize the current action concisely.