Conceptual Guide Deep Agents The runtime behind production deep agents

Deep Agents is a production‑ready harness that builds on LangSmith Deployment (LSD) and the Agent Server to provide durable execution, memory, human‑in‑the‑loop, and observability for long‑running agents. The runtime primitives include durable execution with task queues and checkpointing, short‑term and long‑term memory stored in PostgreSQL, guardrails, middleware, multi‑tenancy, authentication, agent authentication, RBAC, streaming, concurrency control, tracing, sandboxed code execution, MCP, A2A, webhooks, and cron jobs. The Agent Server exposes assistants, threads, runs, memory, and scheduled jobs via a REST API, while LSD manages the underlying infrastructure. Durable execution ensures that agents survive crashes, deploys, and long pauses for human approval by persisting state and resuming from the last checkpoint. Memory is split into short‑term (per thread) and long‑term (per user/organization) to support context across conversations.

The guide maps each production requirement to a runtime capability, illustrating how Deep Agents can be shipped as a package that developers can deploy without reinventing core infrastructure. It emphasizes the need for observable, durable, and multi‑tenant agent systems that can be controlled by operators and users alike.

Key takeaways include the comprehensive set of runtime primitives, the separation of short‑term and long‑term memory, and the use of PostgreSQL for persistence and semantic search.