Gemini Agent: Persistent Work & Multi-Agent Architecture

Google Cloud introduced Gemini Agent on October 8, 2026, bringing questions, knowledge work, content creation, and code execution into one work agent. Instead of choosing an application, model, or workflow first, a user delegates an objective. Gemini Agent plans the work, selects skills and tools, connects to enterprise systems, and returns a finished document, message, analysis, or software artifact.

The consequential change is not another chat surface. Google has separated the model, agent, tools, and execution environment into distinct layers. Gemini Agent can select models at runtime, create subagents, keep jobs running in the cloud, and retain the same context across Gmail, Drive, Docs, Slides, Sheets, Chat, Calendar, Slack, Microsoft 365, command lines, and third-party applications.

One agent for questions, office work, and code

Gemini Agent is available across the web, mobile devices, Windows and macOS desktops, Google Workspace, Microsoft 365, Slack, and command-line environments. It can also operate headlessly inside third-party applications. These surfaces share memory, skills, permissions, and controls, so changing devices or applications does not create a new agent context.

Google describes a broad operating scope:

• answering questions and retrieving organizational knowledge;
• completing knowledge work in Docs, Sheets, Slides, and email;
• creating images and media;
• writing and executing code;
• responding to schedules and events;
• connecting to MCP servers, enterprise software, databases, and desktop files.

The interface is unified, but execution is not compressed into a single model call. Model choice, tool use, and work orchestration remain responsibilities of the agent layer.

Work continues after the laptop closes

Gemini Agent runs in the cloud. Jobs that take hours or days can continue after the user closes a laptop and remain available from another device. Google describes four memory types: session memory for the active task, semantic memory built from organizational knowledge, procedural memory for how work is performed, and episodic memory of previous execution.

This addresses a common break in long-running agent products: a job should not lose its identity and progress when a local process, browser tab, or chat session ends. Persistent execution also raises the runtime standard. The system needs explicit task state, pause and resume behavior, spending limits, permission boundaries, and audit records—not an agent loop that simply keeps taking turns.

Temporary subagents and persistent coworker agents

Gemini Agent can create temporary, job-specific subagents for complex work. Each has an identity and assigned responsibility, while the primary agent coordinates parallel or sequential steps that may run for hours or days.

Google also introduced persistent coworker agents. These operate more like stable organizational roles, with a dedicated identity, an @agents.company.com email address, persistent storage, and scoped context. They can participate in Chat, document comments, and team workflows under their own identity instead of performing every action through an employee account.

The two designs serve different work. Temporary subagents are suited to one-off decomposition and parallelism; coworker agents support lasting roles, continuing context, and team collaboration. Agent Harness ablation research reaches a compatible conclusion: task-specific subagents can improve results, while unbounded generic delegation increases calls and context overhead. Gemini Agent extends that task boundary into identity, storage, and authorization.

The agent and model are separate layers

Gemini Agent is not permanently tied to one Gemini model. Google explicitly separates the agent from its underlying model. The system can currently orchestrate across Gemini models and Anthropic Claude, with additional private and open models planned. Smart Routing selects models according to quality and cost, while a larger project can combine multiple models.

This matches recent model–harness evaluation: the same model can change substantially across harnesses and tasks, and one harness does not form the best fixed combination for every model and workload. Gemini Agent turns model fit into a runtime concern instead of requiring the user to bind every job to one model in advance.

A serious routing layer should eventually use at least three kinds of signal:

1. task type and difficulty;
2. a model's observed behavior inside the current tool and harness environment;
3. live token, sandbox, and execution cost.

Routing only by “large model for hard work, small model for easy work” does not fully address model–harness interaction.

Skills, tools, and context become infrastructure

Gemini Agent includes registries for tools and skills. Tools connect the agent to Slack, Confluence, Git, Jira, Salesforce, ServiceNow, BigQuery, Postgres, Snowflake, MCP servers, and desktop files. Skills are modular instructions, knowledge, and workflows that teach the agent how to complete multi-step work.

Teams can publish shared tools and skills, while individuals can create personal skills. The agent selects capabilities at runtime instead of placing every instruction and tool definition in every turn. Google combines this layer with agent identities, fine-grained authorization, auditing, Agent Sandbox, Agent Gateway, and real-time spending caps.

The competitive question is therefore broader than model quality. It is whether a system can organize context, tools, skills, identity, cost, and auditability into a runtime that can operate over time.

Gemini Agent and DeepSeek Harness occupy different positions

Gemini Agent is closer to a managed work system already connected to enterprise data and office surfaces. DSH is closer to a harness in which developers can replace models, Profiles, plugins, and agent loops. Both demonstrate that agent performance is increasingly determined by the runtime outside the model.

What this means for the DSH ecosystem

Gemini Agent identifies four areas worth tracking in future DSH product work and research:

1. Persistent work is not an infinite loop. Long tasks need recoverable state, explicit stopping conditions, human takeover, and spending limits.
2. Subagents need identities and responsibilities. Task-specific delegation is more useful than increasing agent count without boundaries.
3. Model routing must understand the harness. Selection should account for tool reliability, recovery behavior, and task cost—not only model benchmarks.
4. Skills and tools need registries, authorization, and audit. Discovery is only the first step; execution rights, accessible data, and cost matter equally.

Gemini Agent packages these capabilities as a unified enterprise product, while DSH offers a more open environment for experimentation and composition. The useful comparison will not be a feature checklist. It will measure completion rate, elapsed time, model calls, tokens, permission requests, and failure recovery on the same jobs.

Related reading

• Model–Harness Fit: 66 Configurations across Agent Tasks
• Agent Harness Components: Tools, Compression, and Subagents
• How DSH, deepagents, and Pi Are Converging
• DeepSeek Harness Agent Teams

References

• Welcome to Gemini at Work 2026: Introducing the Gemini agent
• Finding the Right Fit: Model–Harness Interactions across Agent Tasks
• Beyond the Model: Demystifying Harness Effects in Software Engineering Agents