Examples

Real example goals organized by what they're good for.

Goals that work well

Live market data

• "What is the current price of NVDA?"
• "Compare TSLA, AAPL, and NVDA prices today"
• "What did gold close at yesterday?"
• "Pull SPX, NDX, and DJI"

These typically finish in 1 step (one quote_lookup call) in 5–15 seconds.

Web research

• "Find the latest news about TSMC's Q3 earnings"
• "Summarize the top 3 results for 'Meta AI Reality Labs spending 2025'"
• "What are people saying about the new Claude Opus 4.7 release?"

Usually 2–3 steps: a web_search followed by 1–2 web_fetch calls on the most promising results. 30–60 seconds.

Mixing tools

• "Find the latest Apple earnings news, then look up AAPL's current price"
• "Search the web for what GLD tracks, then look up GLD's price"

Multi-step goals where the agent has to chain a search and a quote lookup.

Project / personal context

• "What's the system prompt for this project?" (uses project_context)
• "Did I save anything about photosynthesis?" (uses artifact_read or memory_recall)
• "What did I write about pricing last week?" (uses memory_recall)

These hit your own data via project / artifact / memory tools.

Simple questions

• "What is the capital of France?"
• "Explain photosynthesis in one paragraph"

These work, but Chat mode is faster and cheaper for them. Agent Mode shines when the answer requires fresh data or your own context.

Goals that don't work well (yet)

Anything requiring external writes

❌ "Send an email to John saying I'll be late" ❌ "Post this to Slack" ❌ "Open a GitHub PR" ❌ "Add it to my calendar"

The agent has no L3 tools. These will fail with a permission error or a confused "I can't do that" answer.

Long-tail multi-step research

❌ "Research the entire history of every transistor design from 1947 to today and produce a comprehensive timeline"

Default cap is 6 steps, absolute ceiling 20. Massive research goals will hit the limit and synthesize from incomplete data.

Authenticated third-party data

❌ "Check my GitHub commits today" ❌ "What's in my email inbox?" ❌ "Look at my Twitter mentions"

web_fetch works on public pages only. Anything behind a login wall fails.

Image or audio understanding

❌ "Describe the image I uploaded" ❌ "Transcribe the audio recording"

The tools are text-in / text-out. For images, use regular Chat mode (which routes to a multimodal model). For audio, the dictation feature.

Very vague goals

⚠️ "Find something interesting" ⚠️ "Tell me about cars" ⚠️ "What's new?"

These work but produce weak answers. The planner has nothing concrete to search for. Specify a topic, a question, or a desired output format.

Tip: front-load the specifics

Better goals tell the agent what kind of answer you want:

• ❌ "Tesla news" → ✅ "Find the top 3 Tesla news headlines from this week, with one-sentence summaries"
• ❌ "GLD" → ✅ "What is the current GLD price and what did it close at yesterday?"
• ❌ "AI announcements" → ✅ "Find news from the past 7 days about Anthropic, OpenAI, or Google AI product launches"