Kimi Work goal mode pushes office agents from chat to long-running execution

Kimi Work goal mode extends agent runtime to 24 hours while connecting Feishu, DingTalk, WPS, and Notion through a new plugin center.

The update moves beyond single-turn chat replies. Agents now receive a goal, then execute across multiple steps without further user input until the task ends or the time limit hits.

The change arrived on June 18, 2026. Moonshot AI, the company behind Kimi, framed the release as a practical step for Chinese knowledge workers who need agents that finish real office sequences rather than suggest next actions. In its official announcement, Moonshot AI stated: “Goal mode lets agents complete full office workflows with persistent plugin access instead of resetting after every reply.” (Moonshot AI Blog).

Task length changes pressure on context tools

Long execution windows expose any gap in persistent work context. An agent that runs 24 hours must read updated documents, check Slack or email status, and adjust plans without resetting memory every few minutes.

Kimi Work goal mode addresses part of this by routing agent actions through live plugins. The plugins pull current data from supported tools instead of relying only on what the user pasted at the start.

This design still leaves open the question of deeper synthesis. Plugins deliver raw updates; they do not automatically connect a past pricing decision recorded in a meeting note with a current spreadsheet cell. For example, when last week’s Notion note approved a $50k budget increase but the live WPS sheet still shows the old figure, the agent receives both values separately and does not flag the mismatch.

Context access versus raw runtime

The real contest is no longer how many hours an agent can stay active. It is whether the agent maintains accurate, cross-source context the entire time.

remio keeps five memory layers that record meetings, documents, and prior decisions without user re-entry. When an agent inside remio receives a goal, it already holds the relevant background instead of calling plugins after the fact.

Consider a Q3 forecast update workflow: remio pulls the prior budget approval from its memory layer in one step and directly edits the WPS forecast. Kimi Work goal mode must first call the Notion plugin for the note, then the WPS plugin for the spreadsheet; any link between the two items depends on the agent reasoning correctly across separate calls.

Kimi Work goal mode improves on earlier chat interfaces by keeping the connection open. It does not yet match the breadth of sources that a personal memory system captures by default.

Plugin center targets daily office flow

The plugin center currently supports seven services. Users select which tools the agent may read or write to, then assign permissions once.

Early users report the agent can open a Notion page, update a WPS table, and post a summary in DingTalk without separate prompts. The sequence still requires the initial goal to name the correct documents and accounts.

One reported case involved pulling Q2 sales numbers from a WPS file, cross-checking project milestones in Feishu, and drafting a status message. The full run took 47 minutes inside the stated limits.

Limits remain visible

Kimi Work goal mode caps single runs at 24 hours. If an agent needs data outside the plugin list, the task pauses or fails. The system logs the missing source but does not search further.

The 24-hour window also assumes stable API connections. Any outage in a connected service ends the run early. Moonshot has not published recovery statistics yet.

These constraints keep long-running execution useful mainly for workflows already contained inside the supported platforms.

What the shift actually tests

The announcement shows office agents can now be given goals instead of prompts. The next measurable signal will be whether teams adopt the mode for multi-hour sequences or continue to break tasks into shorter chats.

Watch monthly active goals reported by Moonshot, the number of plugin-center connections per account, and any third-party benchmarks that compare output accuracy across 4-hour versus 24-hour runs.

If accuracy holds steady as runtime grows, long execution becomes a standard requirement. If accuracy drops, the market will favor agents that carry richer context from the first minute rather than those that simply stay awake longer.

remio already operates on that principle. Its memory system records context continuously, so any goal starts with complete background instead of requiring the user to rebuild it inside each long run.

Users who want to test persistent context against extended runtime can compare both approaches on their own office data. The difference shows up most clearly in tasks that draw on decisions made days or weeks earlier.