Reference Deep Agent

Source: src/langgraph_kit/graphs/reference_deep_agent.py Agent ID: reference-deep-agent

The full-stack general-purpose reference agent that wires every kit feature together. Clone this agent as the starting point for any new domain-specific agent (see coding-agent for the canonical extension pattern).

Architecture

Built on the deepagents framework with the complete langgraph-kit feature stack:

User Message
    │
    ▼
Middleware Stack (11 middleware)
    │
    ▼
deepagents create_deep_agent()
    ├── Prompt Assembly (PromptComposer)
    ├── Tool Registry (standard + memory + UI + HITL)
    ├── Worker Definitions (researcher, implementer, verifier)
    ├── Backend Factory (memories, notes, state)
    └── Command Dispatcher (7 built-in commands)

Prompt Sections

The reference agent registers 5 core prompt sections:

Section ID	Stability	Content
core_identity	STABLE	Agent identity and capabilities
memory_instructions	STABLE	How to use persistent memory
orchestration_instructions	STABLE	Multi-agent delegation patterns
continuation_guidance	STABLE	When to continue vs. stop
ui_interaction	CONDITIONAL	UI artifact and event usage

Worker Definitions

Three specialized workers for task delegation, defined in core/orchestration/workers.py:

Worker	Definition	Description
researcher	RESEARCHER_DEFINITION	Finds information, reads docs, searches code
implementer	IMPLEMENTER_DEFINITION	Writes code, makes changes, builds features
verifier	VERIFIER_DEFINITION	Reviews changes, runs tests, validates output

Pre-composed as GENERAL_WORKERS — see Workers.

Features Included

• Prompt assembly with stable-first caching
• Persistent memory (CRUD + auto-extraction)
• Tool registry with risk levels
• 7 built-in slash commands
• Context pressure monitoring
• Session notebook continuity
• Async task delegation
• Message queue for busy threads
• HITL approval for destructive operations
• UI artifacts, progress, suggestions, citations
• Skill discovery
• MCP tool integration
• Resilience middleware (completion guard, empty turn, tool error)
• Langfuse observability

Build Function

def build_reference_deep_agent(
    checkpointer,
    store,
    *,
    mcp_tools=None,
    recursion_limit=DEFAULT_RECURSION_LIMIT,  # 100
):
    """Build the reference deep agent with all kit features.

    Returns: (compiled_graph, command_dispatcher)
    """

Returns a tuple of the compiled graph and command dispatcher, both registered together.

Recursion Limit

Defaults to DEFAULT_RECURSION_LIMIT (100) — high enough for realistic multi-turn runs with worker delegation and tool loops, where LangGraph's native default of 25 is not. Raise it for long autonomous runs by passing recursion_limit=<n> to the builder, or override per-invocation with config={"recursion_limit": <n>}. See the agents overview for details.

Middleware Stack

The reference agent uses the full standard middleware stack (11 middleware), built by core.graph_builder.build_middleware_stack():

1. CommandMiddleware
2. RuntimeStateMiddleware
3. QueuedInputMiddleware
4. ToolErrorMiddleware
5. PressureMiddleware
6. ResultPersistenceMiddleware
7. ExtractionMiddleware
8. EmptyTurnMiddleware
9. CompletionGuardMiddleware
10. StopHooksMiddleware
11. PostRunBackstopMiddleware