Capsule React is a ReAct loop coordinator that orchestrates the reasoning-and-action cycle for an AI agent through a stateless state machine.
The tool solves the problem of coordinating multiple specialized components in an agentic system. Rather than embedding orchestration logic within a single monolithic service, Capsule React acts as a pure control flow engine that drives the cycle of fetching context, reasoning, acting, and observing. It coordinates five separate capsules over an IPC event bus: Session for conversation history, Identity for system prompt construction, Prompt Builder for merging plugin contributions, a Provider for LLM streaming, and Tool Router for dispatching and collecting tool results. The tool itself contains no inference logic, prompt building, tool execution, or message storage—it only manages the flow between these specialized components.
Adoption makes sense for teams building agentic systems within the Astrid OS framework who want to experiment with different orchestration strategies without rewriting supporting infrastructure. Because the tool is purely a control flow coordinator, swapping it out for alternative approaches like debate loops, Monte Carlo tree search, or chain-of-verification strategies changes agent behavior without touching any other component. This modularity is valuable for research and iterative development of reasoning patterns. The tool maintains ephemeral per-turn state in a capsule KV store for correlation and turn tracking, while delegating conversation history persistence to the Session capsule.
The project maintains a minimal, focused codebase that reflects its narrow responsibility. Development activity shows steady attention to the core orchestration logic without scope creep into adjacent concerns like LLM integration or tool execution.
