Julep

Manages agent state across multiple conversation turns by persisting session data, conversation history, and agent context to a backend store. Each agent instance maintains a unique session ID that tracks all interactions, allowing agents to recall previous exchanges and maintain continuity without re-prompting. Uses server-side session storage with automatic serialization of conversation state, enabling long-running agents that survive application restarts.

Enables agents to invoke external tools and APIs through a schema-based function registry that maps tool definitions to callable endpoints. Agents receive tool schemas at runtime, generate appropriate function calls based on task requirements, and execute them through Julep's orchestration layer. Supports both synchronous and asynchronous tool execution with automatic parameter binding, error handling, and result injection back into the agent context.

Deploys agents as serverless functions that scale automatically based on demand. Agents are invoked via API calls that trigger execution in isolated containers or functions. The platform handles infrastructure management, auto-scaling, and resource allocation. Supports both on-demand and scheduled execution patterns.

Provides a declarative workflow system where agents execute predefined sequences of steps (prompts, tool calls, conditionals, loops) with state passing between steps. Each step can depend on outputs from previous steps, enabling complex multi-stage agent behaviors. The execution engine handles step scheduling, error recovery, and state transitions, with support for branching logic and iterative loops based on agent decisions or external conditions.

Abstracts away provider-specific differences (OpenAI, Anthropic, Ollama, etc.) behind a unified agent interface, allowing agents to switch between LLM providers without code changes. Handles provider-specific features (function calling formats, token counting, streaming) transparently, with automatic request/response translation. Supports both cloud-hosted and self-hosted models through a consistent API.

Automatically manages conversation history by storing and retrieving relevant past messages for agent context. Implements intelligent context windowing that selects the most relevant conversation segments based on relevance scoring or recency, preventing context overflow while preserving important information. Supports both full history retrieval and summarization-based context compression for long conversations.

Exposes agents through a REST API that enables programmatic agent invocation, message submission, and session management without requiring direct SDK integration. Agents are deployed as stateless services that handle concurrent requests, with session state managed server-side. Supports both synchronous request/response and asynchronous execution patterns with webhooks for long-running operations.

Provides comprehensive logging and monitoring of agent execution, including step-by-step traces, tool call logs, LLM prompt/completion pairs, and error tracking. Execution traces are stored server-side and queryable through the API, enabling debugging, auditing, and performance analysis. Supports structured logging with metadata (timestamps, latency, token usage) for each execution step.

Provides APIs for creating and managing multiple agents and users within a single Julep account, with role-based access control and isolation between tenants. Each agent can be associated with specific users, and session data is isolated by user/agent combination. Supports bulk agent creation and configuration management through API.

Enables agents to access external knowledge bases (documents, FAQs, databases) through semantic search, retrieving relevant information to inform responses. Supports uploading documents, automatic chunking and embedding, and vector-based similarity search. Retrieved context is automatically injected into agent prompts, enabling agents to answer questions grounded in specific knowledge sources.

Enables agents to perform multi-step reasoning by generating intermediate thoughts and decisions before producing final outputs. Agents can be configured to show reasoning steps, enabling transparency into decision-making processes. Supports both implicit reasoning (through prompting) and explicit reasoning steps (through workflow configuration).