Agent Research is an experimental/advanced workflow for assisted literature exploration, hypothesis generation, and verification-oriented research automation. It is not a claim that the model can solve open research problems by itself.
What You Get
- structured research runs that can plan searches, inspect retrieved material, and assemble cited notes
- action candidate evaluation and ranking before tool-heavy steps
- an action decision ledger and action outcome records for reproducibility
- proof obligation tools for tracking what still needs evidence
- portfolio review modes for comparing branches or candidate approaches
- formalization planning and attempt records where the workflow is configured
- exported Markdown reports plus metadata for later review
What Is Experimental
The discovery workflow combines deterministic bookkeeping with model-generated reasoning. The ledger, proof obligation graph, and verifier statuses are structured artifacts; the model's hypotheses, summaries, and branch proposals still need expert review.
Formalization planning is not theorem certification. A report can describe a planned or attempted formalization, but only verified artifacts should be treated as formal evidence.
Where To Click
| Entry point | What happens | | --- | --- | | Agent Research card on Help/Landing | Opens the AI assistant on the research-agent path | | Agent/discovery mode inside the assistant or orchestrator | Starts the same workflow with the selected backend and tools | | Citation or report links | Jump back into source papers, saved library entries, PDFs, or report artifacts where available |
Workflow
1. Describe the research goal and constraints. 2. The agent proposes candidate actions or branches. 3. Tool-heavy actions pass through candidate evaluation and a recorded decision. 4. The run collects sources, citations, proof obligations, and outcomes. 5. Review the report, citations, and verifier state before moving the result into Writer or a publication workflow.
Notes
Agent research can consume more tokens and time than a single assistant question. Use it when you need a traceable multi-step workflow, not when a quick explanation is enough.