open_deep_research

Facts

Openness

Open source

Licence

MIT

Language

Python

Pinned version

1b7d2e8 (2026-08-10)

Topics checked

29 of 33

Kind of agent

Deep research

Source repository (opens another site)

What it does notably well

• Compaction

  By default src/open_deep_research summarizes each fetched web page with a cheaper model (summarization_model=openai:gpt-4.1-mini) and then compresses each researcher's findings…

• Concurrency

  src/open_deep_research is async throughout and uses asyncio.gather to run researcher sub-agents, a researcher's tool calls, Tavily queries and page summaries concurrently.

• Errors and failover

  src/open_deep_research wraps model calls in with_retry, runs tools through execute_tool_safely, which returns the exception text to the model as a string, and degrades to a…

• Eval harness

  open_deep_research evaluates whole research tasks on the Deep Research Bench dataset with six LLM-judge evaluators and publishes per-configuration results in its README…

• Safe repeats

  In src/open_deep_research, tavily_search collapses the results of all its queries into one dict keyed by URL, keeping the first hit and its query, so a page returned by several…

• MCP and auth

  With mcp_config.auth_required set, src/open_deep_research posts an OAuth token-exchange grant (urn:ietf:params:oauth:grant-type:token-exchange, the RFC 8693 type) with the…