Mirascope
Mirascope provides a Python framework for building production-ready LLM agents through declarative, type-safe tool definitions and automatic agent loop execution. The framework integrates versioning, tracing, and cost tracking directly into the development workflow, capturing full execution context for reproducibility. It supports prompt chaining, dynamic routing, and centralized orchestration patterns for complex multi-step AI workflows.
- Artificial Intelligence
- Developer Tools
- AI Agents
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
Developers building LLM applications struggle with the complexity of creating reliable, multi-step AI workflows. Traditional single-prompt approaches fail on complex tasks, lack mid-process validation, and don't handle the stateless nature of large language models, making it difficult to build production-grade agents that can use tools, maintain context, and be debugged effectively.
Solution
Agent Loop provides a Python framework that simplifies building production-ready LLM agents through a declarative, tool-based approach. Developers define tools as typed Python functions and use decorators to automatically handle LLM call execution, tool-call loops, and response resumption. The framework integrates versioning, tracing, and cost tracking into the development workflow, capturing full execution context including prompts, model settings, and input arguments to ensure reproducibility. It supports prompt chaining, dynamic routing, and centralized orchestration patterns, enabling developers to build complex agentic systems where a central planner delegates tasks to specialized executors while maintaining state and handling retries.
Target Audience
Python developers and engineering teams building LLM-powered applications, agents, and complex AI workflows who need structured, observable, and maintainable prompt orchestration. The framework serves both individual developers and teams requiring collaborative prompt engineering and testing capabilities.
Features
- Type-safe tool definitions via `@llm.tool` decorator that generate schema from Python type hints
- Automatic agent loop execution that iterates on tool calls and resumes responses until completion
- Built-in versioning and tracing through the `@ops.version()` decorator to track prompt changes and execution context
- Cost tracking and metadata capture for every LLM API request, including input and output tokens and model metadata
- Support for prompt chaining to break complex tasks into sequential, verifiable steps
- Dynamic routing using LLM-as-classifier to direct inputs to appropriate workflows or models
- Centralized orchestration pattern where a planner component decomposes tasks and delegates to function-specific executors
- OpenTelemetry-based instrumentation for API-level call footprint tracing