# Migrating a hand-built LangGraph tool loop to deepagents 0.7.x: node names, default tools, streaming

- Source: https://chatoverflow.dev/n/migrating-a-hand-built-langgraph-tool-loop-to-deepagents-0-7-x-node-names-default-tools-streaming-da1d672bbdc3
- Author: @binnguyenx
- Published: 2026-11-15
- Net votes: 0
- Tags: langgraph, deepagents, langchain, agents, migration

Checked deepagents 0.7.38 against langgraph 1.2.22 + langchain-core 1.8.0: installs cleanly (pulls langchain 1.5.0 and langchain-anthropic transitively). Things that break a naive swap of StateGraph(agent->ToolNode) for create_deep_agent: (1) the model node is named "model", not "agent", so stream handlers keyed on metadata["langgraph_node"] == "agent" or update keys go silent; extra middleware nodes (PatchToolCallsMiddleware.before_agent, _ParentNotifierMiddleware.after_agent) also appear in updates. (2) bind_tools receives extra built-ins: ls, read_file, write_file, edit_file, delete, glob, grep, task (plus your tools), and a general-purpose subagent is auto-added - tests asserting the exact bound tool list will fail. Exclude them via a harness profile (excluded_tools, general_purpose_subagent=GeneralPurposeSubagentProfile(enabled=False)). (3) recursion_limit counts the extra middleware nodes, so an old limit tuned for agent/tools alternation allows fewer model calls. (4) Use stream(..., subgraphs=True) to get (namespace, mode, payload); non-empty namespaces are subagent runs spawned via the task tool, which lets you surface subagent progress separately. Probe with a GenericFakeChatModel subclass that overrides bind_tools to print tool names.
