The agent

01In Settings, Agent, pick a provider. Use on-device Apple Intelligence when available, connect a hosted provider, or point it at Ollama or LM Studio running locally.Hosted options include OpenAI, Anthropic, Gemini, and OpenRouter, plus OpenAI-compatible custom URLs. Provider keys go to Keychain and never into run records or iCloud.

02Open the panel and ask something about the page you are on.⇧⌘ASettings decides which side it docks to and whether it narrows the page or floats over it.

03Press the sparkle button in the panel to switch the field into AI Search, then describe what you are looking for instead of quoting it.Semantic fuzzy search of the current page. The field's placeholder changes to tell you which mode you are in.

04Turn on loading more of the page before answering, then ask about something far down an infinite feed.It scrolls the page behind a frozen view, then puts you back exactly where you were. Bounded, so endless feeds give up after a few seconds.

05Read the hard limits and set them low on purpose.Turns, tool calls, wall-clock minutes, open pages, model result size, downloads, artifacts, and optional token or cost caps. A finite cost cap stops a run safely even when pricing is unknown.

06Ask for something that needs approval and watch it stop and ask.

07Point Cowork at a folder and ask for a file change.Changes are staged as a transaction you preview and approve. Replacements, moves, and deletions cannot commit without it, and a committed transaction can be rolled back.

08Ask the agent to find a fact in research you captured earlier.The built-in research search reads captured pages, video transcripts, paper notes, and workspace documents across your local ledger, then returns verbatim passages with source URLs.

09Add one App Integration under Automation & Records.Connect a Streamable HTTP MCP server with a bearer token or OAuth, then approve the current identity and tool list. A changed server identity or schema has to earn trust again.

10Ask one run to give two independent comparisons to child runs and combine them.Delegation needs approval. Children receive smaller pieces of the parent's page, origin, browser-session, time, and cost authority, so they cannot grant themselves more access or budget.

11Open the agent timeline and replay window.Conversations, runs, steps, approvals, usage, artifacts, and replay evidence in one local store. Retention is yours to set, down to not retaining anything.

12Open Observability & Page Signals and turn on console events for active runs.Navigation, console, dialog, download, and page-lifecycle signals appear in the local record. Diagnostic text is a separate switch and starts off. The same section can preview a redacted diagnostic bundle before export.

13Look at scoped memory, which is off by default.Each fact has provenance, sensitivity, an expiry, and a scope. Retrieved memory reaches the model as untrusted context and can never grant a tool or a permission.

14Create a scheduled task, then read why it is unavailable if it is.A task stays unavailable until its provider, integration, folder, session, and policy requirements are all satisfied. The review tells you the exact missing piece.

15Turn Show AI features off and look around the app.The panel, the shortcut, the shortcut list entry, and AI Search are all gone. On-device Apple Intelligence touches like visual tab names have their own separate switches.

Things to try

Worth filming