Nous Research has updated its Hermes AI agent to v0.18, introducing a stable Mixture-of-Agents orchestration mode and new workflow flags.

Mixture-of-Agents mode

In MoA mode a committee of top closed or open-source models collaborates on responses rather than a single model producing output.

According to Nous tests, the combined reply from opus-4.8 and gpt-5.5 outperformed each model when evaluated separately.

Reasoning traces from contributing models such as Claude, GPT and Grok now carry explicit labels. That lets users assess each model’s individual contribution to the aggregator’s final answer.

Flags and learning features

• /goal — The /goal flag sets a project objective and requires the agent to verify completion rather than only reporting task termination.
• /learn — With /learn the agent ingests a folder, link or workflow, extracts a skill from that material, and memorizes it.
• /journey — The /journey flag reveals a timeline of learned skills obtained via /learn, which can be edited to remove unwanted items.

Orchestration, projects, and efficiency

Users can delegate batches of tasks to multiple sub-agents concurrently; they return results without blocking the main chat flow.

A new Projects sidebar provides repository-style codebase navigation and branch-like management for agent-driven development workflows.

Self-supervised learning costs decrease by offloading chat analysis to an auxiliary lightweight model that relies on compact summaries rather than rereading full conversations.

Nous notes additional improvements and unspecified refinements in Hermes v0.18 beyond those described above.