Deep Agents

Deep Agents is a open-source LangChain framework for agents that plan, manage files and delegate. GitHub is the vendor source; hosting and model costs depend on the stack.

In practical use, it is strongest for research agents, coding assistants, multi-agent prototypes. Its verified capabilities include task planning, filesystem context, subagents, langgraph. This makes it relevant to builders and business teams seeking a concrete workflow result rather than a general chat window. A useful evaluation starts with one bounded process, representative data, and a measurable outcome such as time saved, completion rate, review effort, learning progress, or pipeline quality. Buyers should test permissions, exports, auditability, failure recovery, and the amount of human review needed before relying on automated output.

Observed pricing was Open source: usage separate. Prices and included usage can change; taxes, model consumption, implementation, hardware, data, or overages may be separate. At least one important detail was not fully verifiable from static HTML, so manual verification is required before publication.

MCP compatibility is prominent: It consumes MCP tools through LangChain MCP adapters. Teams should pilot with real systems and edge cases, checking integration depth, security controls, data retention, regional requirements, and how easily a person can inspect or override an action. They should compare total operating cost, setup effort, administrator workload, support, and change management—not merely the headline feature list. A controlled trial will show whether the product fits an individual workflow, a governed department rollout, or a developer-operated production system.

Verified product detail

Deep Agents is an opinionated open-source harness for long-running agents. Its repository documents planning, filesystem context management, isolated subagents, memory, human approval, reusable skills, and any tool-calling model. LangGraph provides streaming, persistence, and checkpointing underneath. Unlike a managed agent seat, it gives builders control over the model and runtime while leaving authentication, deployment, evaluation, and operations to their engineering team.

Key features to test

1. Hierarchical task planning and progress tracking. Verify behavior, limits, permissions, and plan entitlement with representative data and failure cases.
2. Filesystem tools for context and artifacts. Verify behavior, limits, permissions, and plan entitlement with representative data and failure cases.
3. Subagents with isolated instructions and context. Verify behavior, limits, permissions, and plan entitlement with representative data and failure cases.
4. LangGraph persistence, streaming, and checkpointing. Verify behavior, limits, permissions, and plan entitlement with representative data and failure cases.
5. Human approval, memory, skills, and tool-calling model support. Verify behavior, limits, permissions, and plan entitlement with representative data and failure cases.

Current pricing and total cost

Open-source framework with no standalone subscription price; model APIs, hosting, storage, tracing, and deployment are separate costs. Prices and entitlements can change; include taxes, overages, implementation, support, and review time in total cost.

Practical evaluation

Test a five-step research or coding job with a 20-document input, one failing tool, and a forced restart. Check whether large artifacts move to files, subagents receive narrow instructions, checkpoints resume correctly, and a person can reject tool calls. Record tokens, tool calls, elapsed time, retries, review minutes, and cost per accepted result. Open source does not mean zero operating cost.

Honest strengths and limitations

• Model and deployment flexibility
• Filesystem and subagent patterns help control long-task context
• LangGraph supplies durable execution primitives

• Hosting and model usage are not bundled
• Production safeguards require engineering
• Results depend heavily on tools, prompts, and model choice

• Building multi-step research agents
• Creating coding assistants with delegated repository tasks
• Prototyping durable LangGraph workflows

Alternatives and buying checklist

Compare LangGraph, LangChain, Microsoft AutoGen, Aider with identical inputs, acceptance criteria, security constraints, and budgets. Confirm billing interval, cancellation, retention, export, training policy, regional hosting, uptime, and support. For agents that can write code or act in connected systems, start read-only, grant least privilege, require approval for consequential writes, log every action, and keep a tested rollback path. This review reflects vendor research performed September 14, 2026; select based on accepted outcomes after correction time and operating cost, not the most impressive first prompt.