crewai-ts

Enables creation of specialized AI agents with defined roles, goals, and backstories that collaborate to complete complex tasks through a coordinator pattern. Each agent maintains its own LLM context and can delegate work to other agents or execute tasks independently, with the framework handling message routing, state management, and execution sequencing across the agent network.

Provides a declarative system for registering tools/functions that agents can invoke, using JSON schema definitions to enable LLM-native function calling across multiple provider APIs (OpenAI, Anthropic, Ollama). The framework handles schema validation, parameter marshalling, and error handling, allowing agents to autonomously decide when and how to use tools based on task context.

Provides full TypeScript support with type definitions for agents, tasks, tools, and configurations, enabling compile-time type checking and IDE autocompletion. Type safety extends to tool schemas, output validation, and callback signatures, reducing runtime errors and improving developer experience.

Abstracts LLM interactions behind a unified interface that supports multiple providers (OpenAI, Anthropic, Ollama, and compatible APIs) and models, handling authentication, request formatting, response parsing, and error handling transparently. Agents can switch between models or providers without code changes, enabling cost optimization and model experimentation.

Provides a task abstraction that encapsulates work units with descriptions, expected outputs, and assigned agents, supporting both sequential execution (tasks run one after another with output chaining) and parallel execution patterns. The framework manages task state, input/output mapping, and dependency resolution, allowing complex workflows to be defined declaratively.

Manages conversation history and context state for agents, maintaining message logs, agent-specific memory, and shared context across task execution. The framework provides hooks for custom memory backends, enabling integration with external storage (databases, vector stores) while maintaining in-memory caches for performance.

Automatically parses and validates LLM outputs against expected schemas, converting raw text responses into structured data (JSON, objects) with type checking and error recovery. Supports multiple output formats and provides fallback strategies when parsing fails, ensuring downstream code receives validated data structures.

Provides a callback/event system that fires at key execution points (agent start, tool call, task completion, error) allowing external monitoring, logging, and custom behavior injection. Callbacks receive structured event data and can modify execution flow or trigger side effects without modifying core agent code.

Provides a declarative configuration system for defining crews (collections of agents and tasks) with settings for execution mode, logging, memory backends, and LLM provider defaults. Configuration can be loaded from code, environment variables, or files, enabling environment-specific behavior without code changes.

Implements a manager-agent pattern where a designated manager agent coordinates work among subordinate agents, making decisions about task delegation, prioritization, and result aggregation. The manager uses the same LLM-based reasoning as other agents but operates at a higher level of abstraction, enabling complex multi-agent workflows without explicit orchestration rules.

Provides built-in error handling for LLM API failures, tool execution errors, and parsing failures, with configurable retry strategies including exponential backoff, jitter, and maximum retry limits. Errors are categorized (transient vs. permanent) to enable intelligent retry decisions without manual intervention.