First agent

The tools in an agent, called directly and then chosen by the model.

At the end of this page an agent has routed a support message with route, and a chat turn has let the model pick a Jev tool on its own.

Hand over the tools

from strands import Agent
from strands_jev import ALL_TOOLS

agent = Agent(tools=ALL_TOOLS)

ALL_TOOLS is every tool in the package, 19 at this commit, in the three groups the Tools reference lists. Hand over a subset when the agent should only do one thing: Agent(tools=[route, fan_out]). Each tool resolves its endpoint in the same order: a Jev in invocation_state["jev"] on the call, then whatever strands_jev.configure(...) was given, then a Jev() built from the key on the machine.

Route a request directly

result = agent.tool.route(
    message="Third time writing. Webhooks silently drop every event since Tuesday and nobody answers.",
    intents={
        "bug_report": "something that used to work is broken",
        "billing": "charges, invoices, refunds, payouts",
        "complaint": "the writer is unhappy with how they were treated",
        "question": "the writer wants to know something",
    },
    handlers={"bug_report": "engineering", "billing": "finance", "complaint": "human", "question": "code"},
    complex_intents=["complaint"],
)
print(result["content"][1]["text"])

The request carries one Choice over the four intents plus none_of_these and a three-level complexity Score, in one call. The result has intent, confidence, handler, escalate and reason; the intent must clear min_confidence (0.6 by default) or its own entry in thresholds, and the intents in complex_intents also go to a person when the complexity score says so. Fourteen such messages routed 14/14 on 2026-09-29 against a keyword baseline's 11/14; the page Measured has the rows.

Let the model choose

agent(
    "Here are two support messages. Which team should take each one, and which is more urgent?\n"
    "1. 'Charged twice this month, please refund one.'\n"
    "2. 'Since the update the export button does nothing.'"
)

The chat model reads the tool descriptions, which are the docstrings on the Tools pages, and calls route or jev_ask with questions it writes itself. It then answers in prose with Jev's numbers in hand. The default model for Agent() is Amazon Bedrock; any Strands model provider works, the Jev call is the same.

Bring your own client

from strands_jev import Jev

jev = Jev(model="jev-1.13.0")
agent = Agent(tools=ALL_TOOLS)
agent.tool.jev_usage(invocation_state={"jev": jev})

Passing a Jev in invocation_state pins the model id, lets you point at another base_url, and keeps the usage tally per client. strands_jev.configure(jev) does the same for every call in the process.

Edit page