How do you implement a LangChain agent with tool prioritization?

What you need to know

Soft priority versus hard priority

Use soft priority for preferences ("usually try X first") and hard priority for rules that must hold ("never search the web before the internal KB", "no refund tool for unverified users").

Soft: prompt and descriptions

1SYSTEM = """Answer support questions for our broadband service.2Tool order:31. search_help_centre — always try this first.42. get_account — only for questions about this customer's own plan or bills.53. web_search — only if the help centre has no answer, and never for prices."""

Also write each docstring to point at the others: web_search's says "Use only after search_help_centre returned nothing relevant."

Hard: middleware that gates tools

1from langchain.agents import create_agent2from langchain.agents.middleware import wrap_model_call34@wrap_model_call5def help_centre_first(request, handler):6    used = {m.name for m in request.state["messages"] if m.type == "tool"}7    if "search_help_centre" not in used:8        # First step: only the help centre is available, and it must be called.9        kb = [t for t in request.tools if t.name == "search_help_centre"]10        request = request.override(tools=kb, tool_choice="search_help_centre")11    elif not request.runtime.context.verified:12        request = request.override(13            tools=[t for t in request.tools if t.name != "get_account"])14    return handler(request)1516agent = create_agent(model, tools=[search_help_centre, get_account, web_search],17                     system_prompt=SYSTEM, middleware=[help_centre_first],18                     context_schema=SupportContext)

The model cannot choose a tool it cannot see. On the first step it sees only search_help_centre and is forced to call it. After that it sees the other tools, except get_account when the user is not verified. request.runtime.context holds values your backend passes at invoke time, such as context=SupportContext(verified=False).

Other options

• Route before the agent. A cheap classifier chooses a tool group ("billing" or "technical") and the agent only sees that group.
• LLMToolSelectorMiddleware with always_include=["search_help_centre"] keeps the priority tool present while a small model picks the rest.
• ToolCallLimitMiddleware on expensive tools, such as web_search with run_limit=1, makes them a last resort in practice.

A real-life example

A broadband company's help-centre support bot had three tools. The team noticed the bot answering "How do I reset my router?" with steps from a web search — for a different brand of router — about 8% of the time, even though the help centre had the correct article.

A stronger system-prompt rule brought this down to 3%. The help_centre_first middleware brought it to zero, because the web search tool is not visible until the help centre has been searched. Web search calls fell by 70%, and the cost per conversation fell with them. The eval set now includes 40 questions whose expected first tool is search_help_centre.

Follow-up questions to expect

• "What does tool_choice do?" — It tells the model it must call a tool (a specific one, or any); forcing it on every step would stop the agent from ever answering, so apply it only to the first step.
• "How do you handle permissions?" — Remove the tool for that user in middleware, and also check permissions inside the tool itself; never rely on a prompt instruction.
• "How do you test priorities?" — An eval set with the expected first tool or tool sequence, run on every prompt or model change.