agent-deep-research

Runs long-form research questions asynchronously against a hosted model and returns structured output, polling until the job completes rather than holding a session open. Two details show care: a dry-run flag previews the spend before committing, and local documents can be grounded as context through uploads to stores deleted automatically afterwards. The dependencies are firm. It needs a python package runner binary and one of several provider keys in the environment, and without a key nothing runs. Grounding local files means shipping them to a third party, ruling it out for confidential material.

Use case: Running a long research question in the background

This skill runs deep research through a hosted research agent without depending on a separate command line tool. It starts a long job, keeps state in the workspace, and polls until the work returns, so a question that takes many minutes does not tie up a single fragile call. ## What it does - Launches a research run from a plain-language question and tracks it as a session. - Grounds answers on local files by uploading them to a temporary search store for retrieval. - Previews the cost of a run before committing to it. - Returns structured output and polls adaptively rather than on a fixed schedule. - Exposes a printed capabilities manifest so an agent can inspect what the skill supports. ## When it is the wrong choice It needs a package runner and one of several model api keys, plus network access to the provider, so an offline or key-less host cannot run it at all. Grounding on local documents uploads them to an ephemeral store, which is a real consideration for sensitive material even though the store is temporary. Deep research is also slow and billed, so a quick lookup belongs somewhere cheaper. ```bash uv run {baseDir}/scripts/research.py start "your research question" ``` ## Notes Preview the cost before a long run, so an exploratory question does not quietly become an expensive habit. Upload context only when the answer actually depends on local material, and start the job asynchronously rather than waiting on a single call that may outlast the session.