An MIT-licensed agent from Nous Research that learns reusable skills, verifies coding goals, and works across terminal, desktop, web, voice, and messaging. The v0.20 Herald release turns its most important promise—letting an agent work for a while without turning into an expensive loop—into a much more practical proposition, especially with smaller local models.
Closed learning loop with `/learn` and inspectable memory; verifiable `/goal` contracts; model-agnostic local and hosted providers; tool self-recovery; per-turn context compaction; 90→500 tool iterations; grounded research citations; voice, A2A, webhooks, and Nous-reported cold-start improvement from about 14 seconds to 1.8 seconds.
Hermes still grants unusually broad access to commands, files, browsers, and messaging accounts. Its impressive efficiency claims are maintainer-reported rather than independent task benchmarks, smart approval is not a security boundary, and a powerful agent can still multiply model costs or mistakes when given loose permissions.
