Research Harness
Research discovers relevant papers and web sources, reads selected public content, and saves source-scoped artifacts in a persistent thread. It is designed for investigations you will revisit—not a paragraph you have to trust and reconstruct later.
Evidence
Source-scoped
Every finished artifact is tied to a specific retrieved source and its provenance.
Continuity
Persistent threads
Follow-up prompts build on durable source identity without bloating the model context.
Control
Quoted first
Long-running work checkpoints and pauses at an approved cap instead of surprising you later.
1. From question to durable evidence
Research separates provisional search results from evidence you can inspect. The discovery agent chooses from numbered search results, selected sources are retrieved and validated, and extraction creates one structured artifact per source that succeeds.
Flow
Research lifecycle
Ask a question
↓
Discover papers and web sources
↓
Select candidates explicitly
↓
Retrieve and validate public content
↓
Extract a source-grounded artifact
↓
Persist it in the thread for later work2. Why this is different from “research this”
The result is a collection of inspectable source artifacts, not one unsupported synthesis paragraph.
Threads and artifacts persist across runs, while the model sees only a bounded ledger and explicit references.
Each source can fail independently; completed artifacts survive provider failures, restart, or credit exhaustion.
Discovery, retrieval, extraction, usage authorization, and persistence have separate durable boundaries.