AI Sandbox: Live Web Research, Run Controls, and Results You Can Actually Read

Shipped August 23-25, 2026.

Four upgrades to the AI Sandbox (the agentic chat in your portal that researches accounts against live signal data):

Live web research

The sandbox agent can now go beyond the Signal API:

• web_search - structured Google results, trimmed to organic hits + answer box. The agent searches to discover, then reads the winners.
• browse_website - fetches any public page as clean markdown, with automatic escalation for hard targets: plain fetch → JavaScript render → premium residential proxies for bot-protected sites.

Ask "how would our signals help [company]?" and the agent reads their site first instead of guessing. Each turn runs under a browsing credit budget, so a runaway research loop can't surprise you.

Run controls

• Stop button - send morphs to stop mid-run; the model loop and any in-flight Signal API calls halt immediately, and partial output is kept.
• Latency chips + stall detector - every tool call shows its timing, and silence is always explained. No more wondering whether a long run is thinking or dead.

Readable results

Tool results now render as typed, dense tables instead of raw JSON: contact searches show Name / Title / Company / Email / Signals columns with expandable per-signal detail; company and intent results get the same treatment. An 80KB JSON blob is now a table you can scan.

Better starting points

A 48-card starter library (8 per category) rotates on the home screen - the most-clicked prompts stay pinned while fresh ones cycle in, and cards you've used stop repeating.

Open the Sandbox and ask it something hard.

Questions? [email protected]