ncode Desktop Automation
Deep research
Deep research is for questions that need many sources: comparing libraries, collecting the current state of a technology, checking what changed in an API. You ask once; a lead agent plans the angles, worker agents search and read, and a reporter writes an answer with its sources. It runs on its own: close the page and the rounds keep going.
Note Research needs a web search engine. Set one up first under Settings → Deep research → Search providers (see Providers, models and web search). Without an enabled engine, every search fails with "web_search is not configured".
Deep research#
A deep research answers one question from the web with its own team of agents: a lead plans the angles, workers search and read pages, and a reporter writes the answer with its sources. It runs in the background, so you can keep working while it does.
It needs at least one search engine with a key (see web search).
| Level | Rounds × agents | Use it for |
|---|---|---|
| Fastest | 1 × 4 | a quick answer in minutes |
| Medium (default) | 2 × 3 | most questions |
| High | 3 × 4 | a wider, deeper sweep |
| Ultra | 4 × 10 | the most thorough answer |
Each research produces a Markdown answer (result.md) and a rendered HTML report. A designed report, a richer HTML page, is built afterwards in the background: by default after every research except Fastest; you can choose after every research, or only when you ask.
Research settings#
| Setting | Default |
|---|---|
| Agents running at once | 10 |
| Pages read per agent | 5 |
| Only pages newer than N days | off |
| Include or exclude domains | none |
| Time limit per agent | 10 minutes, one retry after a timeout |
| A headline after each round | on |
| Lead, worker and reporter models | the defaults, or one each |
Starting a research#
- Click Deep research in the rail.
- Type your question in What do you want to know?
- Pick a level, and optionally tag it with a project or pick a model for this research.
- Click Start research or press ⌘+Return.
Following it#
The list shows every research with its state. A running one shows its rounds and agents live, and a short headline after each round. You can Stop a running research. One that failed or stopped can be retried with Retry, or Retry with… another model; a research cannot be resumed from the middle.
Pin a research to keep it at the top of the list; delete the ones you no longer need.
Reading the result#
When a research is done:
- Open report shows the rendered report in the app, with its sources.
- Download saves it.
- The designed report, when it is built, is a richer page of the same answer.
Every research is also saved as files (result.md, report.html and, if built, designed.html) in its own folder; see Names for where.
Using a research in a conversation#
Type /deep_research in the composer to attach a finished research to your next message, or /deep_research <id> for a specific one. The assistant then works from what the research found.
Research settings#
Settings → Deep research holds the default level, the limits, domain filters, the models of the lead, the workers and the reporter, and when the designed report is built. The table above lists the defaults.