@blade-ai/agent-sdk

Provides a unified agent runtime that abstracts away provider-specific API differences, allowing developers to swap between OpenAI, Anthropic, and other LLM providers without rewriting agent logic. Uses a provider adapter pattern to normalize request/response formats and handle streaming, token counting, and error handling across heterogeneous LLM APIs.

Enables agents to declare available tools via JSON schemas and automatically route LLM-generated function calls to registered handlers with type validation. Implements a registry pattern where tools are defined with input/output schemas, and the SDK handles schema serialization to the LLM, call validation, and error propagation back to the agent loop.

Provides a structured agent loop that manages conversation history, tool call cycles, and state transitions. The SDK maintains a message buffer, tracks tool invocations, and implements a step-by-step execution model where each iteration calls the LLM, validates outputs, executes tools, and appends results back to context for the next iteration.

Supports real-time streaming of LLM responses at the token level, allowing UI applications to display agent reasoning and tool calls as they are generated. Implements provider-specific streaming protocol handlers (Server-Sent Events for OpenAI, event streams for Anthropic) and normalizes them into a unified event stream that applications can consume.

Maintains a conversation history buffer that tracks all messages (user, assistant, tool results) and manages context window constraints. Provides utilities to inspect history, clear old messages, and estimate token usage to prevent exceeding LLM context limits. Implements a simple FIFO eviction policy for older messages when context limits are approached.

Implements automatic retry logic for transient LLM API failures (rate limits, timeouts, temporary outages) using exponential backoff with jitter. Distinguishes between retryable errors (429, 503) and permanent errors (401, 404), and provides hooks for custom error handling and logging. Includes configurable retry budgets to prevent infinite retry loops.

Provides a fluent builder API for configuring agents with LLM provider settings, tool definitions, system instructions, and execution parameters. Uses dependency injection to wire together the LLM client, tool registry, and message history, allowing for easy testing and swapping of components. Configuration is validated at initialization time to catch errors early.

Enables agents to return structured responses (JSON, objects) with schema validation, ensuring that agent outputs conform to expected types. Uses JSON Schema validation to parse and validate LLM-generated JSON, providing type-safe responses in TypeScript. Includes fallback handling for invalid JSON or schema mismatches.