Hermes Agent v0.18.0 Ships with Mixture-of-Agents, Verification System, and Zero Priority Bugs

Hermes Agent v0.18.0, dubbed “The Judgment Release,” shipped July 1st with a milestone achievement: zero critical or high-priority bugs remaining in the entire codebase. The release from NousResearch closed nearly 700 priority issues while introducing major features around model ensembles, work verification, and self-improvement.

The development team spent twelve days in a focused sprint to clear every P0 (critical) and P1 (high priority) issue across the repository. That effort closed 496 issues and merged 196 pull requests, bringing the priority backlog to zero for the first time.

Beyond bug fixes, v0.18.0 transforms how Hermes thinks and validates its work, with new systems for combining multiple AI models and proving tasks are actually complete rather than just claimed finished.

Mixture-of-Agents Becomes First-Class

The biggest user-facing change makes Mixture-of-Agents (MoA) work like any other model selection. Previously a toggle mode, MoA presets now appear as selectable models under an “moa” provider in every interface. Users can pick “my-council” the same way they’d choose Claude or GPT-4, and Hermes automatically routes prompts through the entire ensemble.

When an MoA ensemble runs, each reference model’s complete reasoning displays in labeled blocks before the aggregator synthesizes a final answer. Instead of waiting through silence for a result, users can read what each model in the committee thought, then watch the final response stream live. This transparency works across the command line, terminal interface, and desktop app.

The system also gained proper model identity handling and context window detection, making MoA ensembles as reliable as single-model calls for production use.

Verification System Proves Work Complete

Hermes now records verification evidence for coding tasks and can judge completion by actually running project checks rather than trusting model assertions. The /goal command gained “completion contracts” where users specify what “done” looks like, and the standing goal loop evaluates against that evidence instead of stopping when the model feels satisfied.

A new verification evidence ledger tracks canonical project checks detected by the coding context system. Users can wire in custom verification through a pre_verify hook, and the system includes sensible defaults that distinguish between “I think I fixed it” and “the tests pass, here’s proof.”

The feature ships with verification disabled by default but includes a migration system to tune settings appropriately for different surfaces and use cases.

Self-Improvement Gets Structured

Two new commands formalize how Hermes learns from experience. The /learn command can distill reusable skills from directories, URLs, or workflows users just demonstrated. It writes skills following the standards in a project’s CONTRIBUTING.md file automatically, turning any workflow into a repeatable capability with a single command.

The /journey command provides a timeline view of everything Hermes has learned, displaying accumulated memories and skills with options to edit or delete items directly. The desktop app adds a complementary “memory graph” - a radial timeline visualization that makes the agent’s knowledge base visible and manageable for the first time.

Background self-improvement also became more efficient by routing the post-turn learning analysis to auxiliary models and digesting context instead of replaying full conversations.

Background Task Delegation

The delegate_task function can now fan out multiple subagents that run in parallel without blocking the main chat. Users can kick off research across multiple topics or audit several code modules simultaneously, then continue other work while the fleet operates. When all subagents complete, their results return as a single consolidated summary instead of requiring individual monitoring.

The CLI and terminal interfaces track background subagents in the status bar, providing visibility into ongoing parallel work.

Desktop App Gains Coding Focus

The desktop application added first-class project management with a sidebar of codebases, dedicated coding rail, review pane, and git worktree management. The new project system follows a project → repo → lane model that organizes coding work into structures the agent can understand and act on.

Additional desktop improvements include multi-terminal panels with persistent tabs, PR-style file diffs in chat, and a floating composer window. The interface also gained conversation timeline navigation for long threads and context usage breakdowns.

Production-Ready Gateway

The gateway system can now scale to zero when idle and coordinate clean shutdowns without dropping active conversations. Hosted or relay-only deployments can go dormant when unused and wake on demand, while lifecycle actions like restarts coordinate external drains to avoid cutting off users mid-conversation.

The release also added Google Vertex AI as a first-class provider, handling OAuth2 token refresh automatically so organizations using Gemini through Google Cloud can connect without manual token management.

Bottom Line

Version 0.18.0 represents a maturation milestone for Hermes Agent, moving from experimental features to production-ready systems. The zero-bug achievement demonstrates serious engineering discipline, while the new verification and MoA systems address real limitations in AI agent reliability. The self-improvement formalization makes agent learning visible and controllable rather than mysterious.

For teams already using Hermes, the upgrade brings immediate value through better model ensemble handling and background task management. The verification system, while disabled by default, offers a path toward more trustworthy automated work when enabled thoughtfully.