Hermes Agent 0.18.0: Mixture-of-Agents as a selectable model, evidence-based completion

Major release. Mixture-of-Agents stops being a mode and becomes a selectable model under a moa provider in every picker, with each reference model's output shown as its own block and the aggregator's answer streamed live. The agent now records verification evidence for coding work and judges completion against /goal completion contracts rather than self-assertion, with a pre_verify hook and a one-time defaults migration. Also adds /learn to distil reusable skills and /journey plus a desktop memory graph to inspect and prune what the agent has learned.

PUBLISHED2026-07-01

OBSERVED2026-08-11

AGE1mo

SOURCES1

Mixture-of-Agents

Named MoA presets appear as selectable models under a moa provider across CLI, TUI, desktop and gateway
Each reference model's full output renders as its own labelled block before aggregation
Aggregator answer streams live instead of appearing all at once

Verification and goals

Verification evidence recorded for coding work; completion decided by running project checks, not assertion
/goal gains completion contracts judged against that evidence
pre_verify hook for custom checks; one-time migration tunes defaults

Learning

/learn <anything> distils a reusable skill from a directory, URL or a just-completed workflow, written to the standards in CONTRIBUTING.md
/journey timeline in CLI and TUI shows accumulated memories and skills, editable and deletable
Desktop memory graph: radial timeline of memories and skills

Platform

Gateway scale-to-zero and drain coordination
First-class coding projects in the desktop app
Subagents can fan out in the background

Backlog sweep

All P0 and P1 issues and PRs closed: 496 issues, 196 PRs (~692 items) in the window; ~1,950 total issues and PRs closed
Since v0.17.0: ~1,720 commits, 998 merged PRs, 2,215 files changed, 949 issues closed, 370+ contributors

COMMUNITY