Back to Blog

Engineering

AgentsLocal executionCloud tasks

Local, Cloud, and Personal: how the three agents differ

Log something on your phone, leave a task running in the cloud, or pick up an ongoing conversation. Find the agent that fits.

ISEA Team

Sometimes you want to log an expense on your phone. Other times, you want to hand off a task and check back later. ISEA offers three agents for these different ways of working.

They share record rules and some tools. What differs is where they run, which data they can read, and how the conversation continues.

Local Agent: tools on your device

Local Agent can read and edit records on your phone. The model interprets your request and chooses a tool; the app runs it and shows the result in the conversation.

“Local” describes tool execution and record access. It does not mean the model runs offline on the phone. AI inference currently uses cloud model services. Related inputs and tool results may pass through cloud processing. Turning off record sync does not turn off AI requests.

This is a direct entry point for recording, querying, and editing on your phone. Execution depends on the client running; you should not assume every step continues after you exit the app.

Cloud Agent: tasks on the server

Cloud Agent runs tasks on the server and saves progress, messages, and steps that need your confirmation. Work can continue after your phone disconnects. When you return, you can see the saved results.

It works with the account's available cloud data. A record that exists only on your phone and has not synced is not directly accessible. Cloud writes must also sync back before they become the phone's latest local records.

Service failures, model errors, cancellation, and requests for confirmation can still stop a task. Check its status when you return; the system does not retry every action indefinitely.

Personal Agent: an ongoing conversation

Personal Agent keeps an ongoing main conversation and runs on the server. You can pick up an earlier topic, add details, change a plan, or ask it to handle records and reminders.

You can add “move that reminder to the afternoon.” New input participates at execution boundaries, but stopping or correcting a request cannot undo writes that already completed. Those need a subsequent corrective operation.

Long conversations also have limits. Context can be summarized, and summaries may lose details. Important facts should come from records, and completion from tool results. Lasting preferences are managed through shared memory.

Choosing an entry point

Your taskSuitable entry pointBoundary
Quickly work with records on this phoneLocal AgentThe app executes tools; inference still needs a connection
Leave a cloud-capable task running and check laterCloud AgentOnly data available in the cloud can be accessed
Continue a conversation, add background, and follow upPersonal AgentServer execution, bounded context, and explicit action results

Availability depends on version, account configuration, and permissions. These are not three unrelated business databases. Accessible records follow shared identities and contracts; execution location, sync progress, and conversation organization make the difference.

Knowing when work actually finished

Check the action result in the conversation, or open the record itself. ISEA uses tool results to show what was saved and whether an action succeeded. The AI's written reply alone cannot confirm a save.

File imports also require preparation, validation, and confirmation. If something fails, check tool results and the record screen rather than inferring success from fluent prose.

Next: from conversation to an interface you can use.