Research Assistant

Drop a one-line topic like “recent EV subsidy trends”, “RAG evaluation methods”, or “competitor pricing this quarter” — and the agent goes off, collects sources, balances supportive and skeptical views, and hands you back a tidy canvas document.

What you get

Send a topic in chat and shortly after you’ll see:

One canvas document — a report that opens automatically in the in-app canvas panel. Five or more sources are cited inline as links, and the layout always runs in the order: brief summary → supporting views → opposing views → open questions.
(Optional) download link — if there’s any extra deliverable (a CSV with lots of tables, a PDF of source excerpts), it shows up in chat as a downloadable file.

Ask the same topic again later and the canvas with the same title gets a new version instead of a fresh document — so you can compare how the picture changed over time in one place.

Tools to enable

Open the agent → Settings tab → Enabled Tools and turn on:

Settings label	Why you need it
Web Search	Pulls candidate sources by keyword and keeps the citation links
HTTP Request	For static HTML or JSON APIs that don’t need JS rendering — lighter and faster
Web Browser	For pages that require JS rendering, clicking, or scrolling — SPAs and dynamic dashboards
Canvas Create/Update	Saves the organized report as an in-app canvas; same topic → updated as a new version
Share File	(Optional) only when a separate deliverable file is needed. Asks for your approval before it runs
Notify User	Autonomous runs only (schedules, webhooks, mentions, etc. — when the agent runs without you present). In a regular chat conversation the agent’s text reply already reaches you directly, so this tool is not shown

Example system prompt

Paste this into the agent’s System Prompt field on the Settings tab (tweak to taste):


You are a research assistant. When the user sends a one-line topic, follow this procedure:

1. Break the topic into 2–4 sub-questions and run web searches.
2. Collect at least 5 sources, prioritizing primary sources (official sites, institutions, reputable outlets).
3. Always organize the output into these four sections:
   - Brief summary (one paragraph)
   - Supporting views
   - Opposing or skeptical views
   - 3 open questions that remain unresolved
4. Every claim in the body gets an inline citation in [Title](URL) form.
5. Save the final result as a canvas. Use the user's one-line topic verbatim as the canvas title.
6. After saving, send only one line in chat: "Saved to the canvas."

How it flows

Type the topic: Send a one-liner in chat. For example: RAG evaluation methods 2026.
Source collection: The agent runs several web searches on its own (a single search may also break the topic into multiple sub-questions to explore). You’ll see progress live in chat.
(Optional) deeper read: If snippets aren’t enough, it opens specific pages to back up citations.
Canvas write-up: The organized report opens automatically in the canvas panel on the right.
Wrap-up: A single line — “Saved to the canvas.” — appears in chat.

Send the same topic later and the canvas updates as a new version under the same title instead of creating a duplicate, so you can compare across time.

Tweaks

A one-line tweak in the system prompt covers most preferences.

Domain limits: Add “Prefer .gov, .edu, and reputable outlets as sources.” The web search tool accepts allowed_domains (an allow-list) or blocked_domains (a block-list) — the two are mutually exclusive and cannot be set at the same time.
Length cap: “Keep the canvas body under 600 characters.” gives you a tight briefing instead of a long report.
Citation style: “Use [^1] footnote markers and list sources at the bottom.” switches to an academic style.
Language balance: “Cite at least 2 sources from English outlets and 2 from local outlets.” enforces a multilingual mix.

Advanced

You don’t need this section unless you’re deciding between canvas and share-file, or want repeat research to get smarter over time.

The two tools use different output channels. It’s common to use both in one research run.

Situation	Pick
User should read and edit it in-app	Canvas Create/Update — same title → new version on the next call. Pass `force_new=true` to create a separate new canvas even when the title matches
External sharing, attachments, or archiving	Share File — delivered as a download link in chat
Table-heavy data, needs CSV	Share File (pass CSV text in `content`)
Pre-built binary (PDF, image)	Share File (use `path` to point at the file in the sandbox)

Canvas is visible only inside the user’s chat — not to other agents on the same team. If you need to share an output with teammates, export with share file, or use the team shared state tool (Team State Set) for information the whole team should see.

Memory

If you research the same domain over and over (one industry vertical, recurring competitor list), turning on the memory tools pays off.

Settings label	What it adds
Memory Scan	Recalls earlier topics and key findings so you don’t repeat searches
Read File / Write File / Edit File	Accumulates definitions, glossaries, and trusted source lists in agent-private notes. Relative paths (e.g. `memory/sources.md`) map to the agent’s memory area; absolute paths (e.g. `/tmp/data.csv`) can reach anywhere in the sandbox. The persona file is platform-managed and write-protected

With memory enabled the agent automatically files lessons learned at the end of a job and uses them first on the next call. To pull up results from a past research session, the history read and search tools let you reach archived segments from earlier conversations. Memory is agent-private, so it isn’t visible to teammates. For team-wide reference material, use team state or canvas.

Cost notes

Web search uses credits per provider call, and the LLM also burns credits while summarizing. If you have HTTP Request or Web Browser enabled, reading individual pages adds LLM context cost on top. A typical 5-source run is light overall, but a strong “use 20+ sources” instruction inflates calls and cost — set a sensible ceiling in the prompt.