behalf

A small framework for the recurring pattern of a scoped agent: hand it a brief, a fixed set of tools, and bounded authority, and it does one job.

Abstractions

• Task what to elicit from the user, which tools the agent gets, and what a valid result is.
• ToolSpec can be a single tool or action.
• runner backend is the model like the Claude Agent SDK, Google ADK, or Strands on Bedrock.

from behalf import run_task, make_runner
outcome = await run_task(MyTask(), make_runner("aws"))

Manifests

manifest_schema() is sent to the model during setup (see render_setup_prompt), and finalize_setup only accepts a manifest that has every schema field. run_task raises ValueError if a manifest, conversed or passed in, is missing a field.

Results

Runners return None from run_agent; results come from tool side effects. record_tool(sink, fields) returns a tool that appends each call's arguments to sink:

from behalf import record_tool

class MyTask(Task):
    def build_tools(self, manifest):
        self.results = []
        return [record_tool(self.results, {"job": str, "site": str})]

Install

pip install behalf            # core only (no backend)
pip install behalf[claude]    # + Claude Agent SDK
pip install behalf[aws]       # + Strands / Bedrock
pip install behalf[gemini]    # + Google ADK

License

HPCIC DevTools is distributed under the terms of the MIT license. All new contributions must be made under this license.

See LICENSE, COPYRIGHT, and NOTICE for details.

SPDX-License-Identifier: (MIT)

LLNL-CODE- 842614