Deep Agents Overview

SHIPIT Agent's deep agent system provides autonomous, self-directing agent capabilities that go beyond LangChain. All deep agents support MCP servers, built-in tools, real-time streaming, and memory.

Deep agents also support the same skills system as plain Agent, including packaged skills, custom skill catalogs, explicit skills=[...], and prompt-based auto-matching. See the Skills guide.

Agent	What it does	Key feature
GoalAgent	Decomposes goals, tracks success criteria	Autonomous multi-step execution
ReflectiveAgent	Self-evaluates and revises output	Quality threshold with scores
Supervisor	Delegates to workers, reviews quality	Hierarchical multi-agent
AdaptiveAgent	Creates new tools at runtime	Dynamic tool creation
PersistentAgent	Checkpoint and resume	Long-running tasks
Channel	Typed agent-to-agent messaging	Structured communication
AgentBenchmark	Systematic agent testing	Pass/fail reports

All deep agents support:

# All tools (web search, code exec, etc.)
agent = GoalAgent.with_builtins(llm=llm, goal=goal)

# Real-time streaming with output content
for event in agent.stream():
    print(event.payload.get("output", ""))

# Memory across runs
agent = GoalAgent(llm=llm, memory=memory, goal=goal)

Runnable examples

• python examples/11_deep_agents.py — all deep agents
• Notebooks: 14, 16, 17, 18, 19, 20, 21