Multi-Agent Orchestration in the Desktop

Multi-Agent Workflows in Hermes

Hermes supports multi-agent orchestration: a primary agent that coordinates one or more subagents, each running independently and reporting results back. This is covered in depth in the Hermes Agent Advanced track (T19), but the Desktop app adds a visual layer to orchestration that makes it significantly easier to follow and debug.

When a primary agent spawns a subagent, the subagent gets its own session. In the terminal, tracking multiple concurrent agent sessions requires multiple terminal windows or careful log management. In the Desktop, subagent sessions appear in the session list as they are created, the Agents panel shows their status in real time, and the preview rail can display any session's activity on demand.

This visual layer does not change how orchestration works -- the underlying agent behavior is identical whether you are watching it from the Desktop or a terminal. But it makes orchestration dramatically more accessible by surfacing what is happening across all active sessions without requiring you to manage multiple terminal contexts simultaneously.

The Agents Panel

The Agents panel (accessible from the left sidebar or the View menu) shows all active agent sessions in the current Hermes instance, including both the primary session you are chatting with and any subagents it has spawned.

Each entry in the panel shows: the session name or ID, the agent's current status (idle, working, waiting for approval), the current task or the most recent tool call, and a progress indicator if the agent is in the middle of a multi-step operation. You can click on any entry to bring that session's conversation into focus in the main chat view.

The panel updates in real time. As subagents start, work, and complete their tasks, their entries move through status states. When a subagent finishes and returns its result to the primary agent, it shows as completed. This gives you a live dashboard of your multi-agent workflow's progress without needing to ask any agent for a status report.

The Command Center

For more complex orchestration scenarios -- workflows with many agents, structured task delegation, or agent teams with defined roles -- the Command Center provides a higher-level view. It is a dedicated panel that surfaces the Hermes Command Center feature (if configured in your setup) directly in the Desktop UI.

The Command Center is particularly useful for workflows derived from T21 (Hermes for Homelab) or custom multi-agent setups from T19 (Hermes Agent Advanced). If you have a configured Command Center workflow, the Desktop panel lets you trigger it, monitor it, and inspect individual agent outputs -- all from the same interface you use for interactive chat.

For users new to multi-agent orchestration, the Command Center panel is a good way to experiment with structured workflows. You can trigger a predefined workflow, watch the agents execute it from the Agents panel, and examine the outputs without manually coordinating the orchestration yourself.

Following Parallel Agent Activity

The most common multi-agent pattern in Hermes is parallel execution: a primary agent delegates several independent subtasks to subagents simultaneously, waits for all of them to complete, and then synthesizes the results. In the Desktop, you can follow this at several levels of detail.

At the broadest level, the Agents panel shows all sessions and their statuses. A glance tells you how many subagents are active, which are working, and which have completed. This is the overview view -- useful for answering "are all my agents done yet?"

For more detail on any specific subagent, click its entry in the panel to bring its conversation into the main chat view. You see the full session: what task the primary agent gave it, what tools it called, what it found, and what result it reported back. This is the detail view -- useful for understanding what a specific agent did and why.

The preview rail adds a third layer. If a subagent is working on files, the file browser updates to show changes as they happen. If a subagent is reading web pages, the preview rail can render the pages it is looking at. You do not need to switch between agent sessions to see what each agent is doing -- the preview rail shows you the artifacts they are producing in real time.

Practical Orchestration Use Cases

Multi-agent orchestration in the Desktop is particularly effective for a few workflow types:

Parallel research. A primary agent delegates research on three different topics to three subagents simultaneously. While all three are working, the Agents panel shows their progress. When each finishes, the primary synthesizes the results. The Desktop makes the parallelism visible -- you can watch three conversations happening at once rather than wondering when a single sequential research task will end.

Code review across multiple files. A primary agent sends a subagent to review each of three modules independently, then collects the findings. The Agents panel shows all three review sessions, and clicking on each shows the specific issues found. The file browser tracks which files each reviewer is examining.

Test and fix loops. A primary agent runs tests, delegates individual test failures to subagents that investigate and fix each one independently, then verifies the fixes. The Agents panel shows the fix subagents working in parallel, making it easy to see when all failures have been addressed.

Approval Prompts in Multi-Agent Workflows

Approval prompts (the default confirmation requests before risky tool calls) operate independently per session. If a subagent needs to take a risky action, the approval request appears in that subagent's session view. The primary agent pauses until the approval is granted or denied.

In multi-agent workflows, it is common to enable the YOLO toggle for subagent sessions to avoid interrupting the orchestration flow with approval prompts. Since subagents are typically executing well-defined, scoped tasks rather than open-ended exploration, the risk profile is lower and the approval overhead is less valuable. The Agents panel makes it easy to enable YOLO for specific subagent sessions while keeping the primary session under standard approval rules.