Hermes Agent v0.5.0 drops

Nous Research published Hermes Agent v0.5.0 (tag v2026.3.28) on March 28, 2026 and described the release as a “hardening” update. (newreleases.io)) The release connects Hermes to the Nous Portal with access to 400+ models and adds a first‑class Hugging Face Inference API provider that includes a curated agentic model picker and a live /models probe. (github.com)) Plugin lifecycle hooks — pre_llm_call, post_llm_call, on_session_start, and on_session_end — now fire inside the agent loop and across the CLI/gateway, providing defined hook points for telemetry, audit logging, and plugin-driven observability. (github.com)) Supply‑chain hardening in v0.5.0 removed a compromised litellm dependency, pinned dependency version ranges, regenerated uv.lock with hashes, and added a CI workflow that scans PRs for supply‑chain attack patterns. (github.com)) The runtime plumbing was simplified by replacing a tunneled backend with a native Modal SDK implementation (Sandbox.create.aio + exec.aio), and reproducible deployment support was added via a Nix flake plus a NixOS module with persistent container mode and auto‑generated config keys. (github.com)) Anthropic integrations switched from a hardcoded 16K max_tokens to per‑model native output limits — for example 128K for Opus 4.6 and 64K for Sonnet 4.6 — to eliminate truncated responses and thinking‑budget exhaustion. (github.com)) The changelog lists 50+ security and reliability fixes alongside UX and orchestration features such as Telegram Private Chat Topics and an overhaul of the /model command. (newreleases.io)) Independent coverage of the release highlights added Exa support and small GPT‑5.4 behavior tweaks called out in the published changelog for v0.5.0. (ai-primer.com))