DeepSeek Harness Roadmap Unveils Mobile Access, Persistent Memory & Multi-Agent Teams

A chat screenshot shared by user "Cui Xiaotian" surfaced the DeepSeek Harness team's public product roadmap, highlighting a line that reads "support for remote and mobile use." The roadmap, originally published in the article "DeepSeek Harness Desktop: Ready to Use Out of the Box," lists six forward-looking directions.

Remote and Mobile Access

The most immediately discussed item is remote and mobile access. The article explains the practical driver: when an agent runs a long development task — modifying code, running tests, iterating on results — the process can last hours. Users who step away from their desktops currently have no way to monitor progress or intervene when the agent pauses for confirmation or additional input. Mobile access would let developers check status and send guidance from a phone, reducing device-switching friction. The team has not yet disclosed whether this will be a native app, a web interface, or another form factor, nor a concrete launch timeline.

Personalized Long-Term Memory

Long-running projects span multiple chat sessions. Each new session forces the user to re-explain context: why requirements were set, which approaches were tried, what issues remain open. This repetition wastes time and risks omitting critical details. The roadmap proposes personalized long-term memory so that accumulated background and preferences persist across sessions and continue to inform future tasks. Implementation details — how memory is stored, updated, and user-controlled — are not yet specified.

Browser and GUI Software Automation

Real work rarely stays inside a single window. Research may require opening web pages; data verification may need spreadsheets; development may involve checking a browser-rendered UI or interacting with other desktop applications. The roadmap aims to extend the agent's ability to operate graphical interfaces, letting more GUI-dependent steps enter the automation pipeline. This expansion directly raises the need for stronger execution isolation.

Improved Sandbox Functionality and Security

As the agent gains access to more files, tools, and applications, the execution environment must enforce boundaries: constraining the scope of each run and isolating concurrent tasks. The roadmap treats sandbox hardening as a co-requisite for broader automation capabilities.

Agent Team Collaboration

Complex tasks can be decomposed among multiple agents, but effective division of labor, information passing, and result aggregation remain challenging. The team plans to refine the "agent team" feature — currently delivered as a plugin — to improve both outcome quality and ease of use. The plugin management UI shows agent-team capabilities alongside other extensions.

Session Sharing and Multi-User Collaboration

Distinct from multi-agent coordination, this direction explores how multiple human users can share a session, hand off work, and jointly view or participate in an agent-driven workflow. The team uses the term "explore," indicating that sharing mechanisms, permission models, and product shape are still undecided.

Rollout Strategy and Adoption Signal

The team states it will continuously add opt-in experimental features and promote validated ones to default-on status. Notably, the article cites internal metrics: among DeepSeek API users, DeepSeek Harness ranks first in daily active users and daily session count, making it the most popular coding agent in that ecosystem.