Hermes Agent v0.15: The Velocity Release , What's New in 2026's Biggest Update

Hermes Agent just shipped v0.15.0 ("The Velocity Release") on May 28, 2026 , and then followed up with v0.15.1 (a same-day hotfix) on May 29. This is the biggest single update in the project's history.

747 merged pull requests. 321 community contributors. 282,712 lines of code changed. And one file , the core agent loop , went from 16,083 lines to 3,821.

The highlights:

Sources: Hermes Agent v0.15.0 release notes | Hermes Agent v0.15.1 hotfix notes | Hermes Agent Changelog

The Big Refactor: run_agent.py Goes from 16K to 3.8K Lines

The core agent conversation loop , run_agent.py , was the file nobody wanted to touch. 16,083 lines, 90 seconds to load in your editor, and so dense that even core contributors avoided refactoring it.

v0.15.0 splits it into 14 cohesive modules under agent/. Every extraction keeps a thin forwarder on AIAgent, so behavior doesn't change.

Why you should care:

• Future Hermes development moves faster
• Plugin authors can actually grep the codebase
• The file that took 90 seconds to load in your editor now opens in a blink

Kanban Grows into a Real Multi-Agent Platform

Kanban was already one of Hermes' strongest features. This release turns it into a genuine multi-agent orchestrator, built across 104 PRs.

Feature	What it does
Auto-decomposition	Triage takes one task and breaks it into a tree of sub-tasks automatically
Swarm mode	One command creates a full Swarm v1 graph: root → workers → verifier → synthesizer → shared blackboard
Per-task model overrides	Use cheap models for boilerplate tasks, expensive ones for hard sub-tasks
Scheduled tasks	Tasks can have scheduled start times
Worktree per task	Each task gets its own Git worktree with its own branch
Retry fingerprinting	Failed tasks are detected and retried without duplicating work
Stale task detection	Workers that stop reporting get re-assigned
Drag-to-delete trash zone	Clean up stale tasks visually

For anyone running multi-step CI pipelines, parallel code reviews, or model evaluation workflows , this is the feature to explore.

session_search Rebuilt: Instant, Free, No LLM

The old session_search was an aux-LLM-powered tool. It cost about $0.30 per call and took ~30 seconds to find and summarize three past sessions. Sometimes it invented results when the right session wasn't in the hit list.

The new session_search is a single tool with three modes (discovery, scroll, browse) inferred from which arguments you pass. No aux-LLM, no configuration knob, no companion skill.

Discovery: ~20ms instead of ~90s (4,500x faster) Scroll: ~1ms

Searching your past sessions for context is now instant and free. This is the kind of quality-of-life improvement that changes how you use the agent day to day.

Cold-Start Performance: Another Second Saved

Three optimization rounds in v0.15.0:

• Deferred imports: openai._base_client loads 240ms later and uses 17MB less memory
• 47% fewer per-turn function calls: 399k → 213k for a 31-turn chat
• Adaptive subprocess polling: -195ms per tool call, 1+ second per turn
• Termux cold start: 2.9s → 0.8s
• hermes --version: 701ms → 258ms (63% faster)

This flips the head-to-head benchmark against Codex CLI from 5/11 wins to 6/11.

Promptware Defense: Brainworm-Class Attacks Blocked

Inspired by recent Brainworm / Promptware Kill Chain research, Hermes now defends the context window against three attack vectors:

1. Single source of truth (tools/threat_patterns.py) , ~15 new Brainworm/C2 patterns flagged as malicious content
2. Recalled memory scanned at load time , memory snapshots are checked for injection payloads before they enter context
3. Tool result delimiter markers , a malicious file or remote service can't impersonate Hermes' own system content

Plus a new security-guidance plugin that pattern-matches dangerous code writes before the agent executes them.

Bitwarden Secrets Manager: One Token to Replace All API Keys

If you've ever managed a dozen API keys in ~/.hermes/.env, this is for you.

Hermes now auto-installs Bitwarden Secrets Manager (bws) on first use. Point it at Bitwarden with one bootstrap token (BWS_ACCESS_TOKEN), and every credential , OpenAI, Anthropic, xAI, Discord, Telegram, email , comes from Bitwarden at startup.

Key details:

• Rotate a key in the Bitwarden web app → Hermes picks it up on next restart
• EU Cloud and self-hosted Bitwarden server URLs supported
• Detected credentials are labeled with their source so you can see which came from Bitwarden vs local config

ntfy: Push Notifications Without an Account (23rd Platform)

ntfy joins as the 23rd messaging platform , a self-hostable push-notification service with no signup, no API key, just a topic URL.

Your agent can now send you push notifications from:

• Cron jobs completing
• Kanban task finishing
• Any send_message call in a script

To your phone, your watch, your desktop, your homelab. Zero account required.

Skill Bundles: One Slash Command Loads a Whole Workflow

A skill bundle is a named group of skills that loads them all together with one slash command.

Example: set up a "writing day" bundle with humanizer + ideation + obsidian + youtube-content, and /writing-day activates all four for the session.

Three new optional skills also land:

• code-wiki , Karpathy's LLM-Wiki pattern for persistent indexed dev wiki
• openhands , Delegate tasks to OpenHands for parallel coding agents
• web-pentest , OWASP-style web pentest recipes

TUI Session Orchestrator

The Ink-based TUI now supports multiple live sessions in one window. List, switch between, refresh, and close sessions without leaving the TUI. Dispatch a new session with a session-scoped model picker.

Plus CJK / IME input rendering improvements, mouse-tracking, scrollback preservation, and x.com link rendering.

Other Highlights

Feature	Summary
Two new image_gen providers	Krea 2 Medium ($0.03) + Large ($0.06), FAL ported to plugin
Nous-approved MCP catalog	Interactive picker for trusted MCP servers
xAI integration round	Web Search plugin, hermes proxy Grok upstream, retired model detection + migration
TTS speech-tag pauses	Natural pauses in AI voice output
Full skills.sh catalog	858 → 19,932 entries surfaced in the picker

v0.15.1 Hotfix (Same Day)

v0.15.0 shipped, and within hours the team pushed 0.15.1 with critical fixes:

• Dashboard 401 reload loop fixed , In loopback mode, the dashboard was caught in an infinite reload. Fixed by the same PR that ships.
• Docker --insecure now explicit , No longer inferred from bind host. Set HERMES_DASHBOARD_INSECURE=1 explicitly if needed.
• MCP bare command resolution , npx/npm/node now resolve in Docker containers
• SIGTERM kills Kanban workers , Workers actually terminate when sent SIGTERM
• Skills page sidebar restored , Source pills and category filter work again

Should You Upgrade?

Upgrade if:

• You use Hermes daily , session_search and cold-start improvements alone are worth it
• You run multi-step automation , the Kanban platform is the strongest reason to upgrade
• You manage multiple API keys , Bitwarden integration saves real time
• You use Docker or run Hermes headlessly , hotfixes fix significant issues

Probably 0.15.1 not 0.15.0:

• The 401 reload loop in 0.15.0 affects Docker and hosted Hermes setups. Install 0.15.1 directly.

Install:

pip install hermes-agent

Or if you're upgrading:

pip install --upgrade hermes-agent

Source: Hermes Agent GitHub releases | Official docs pinned to the session prompt.