Hermes Agent V2.0 (Refreshed!): This NEW UPDATE to HERMES IS CRAZY!
By AICodeKing
Key Concepts
- Hermes Agent: An evolving agentic framework transitioning from simple chat-based interactions to a persistent, multi-agent coordination system.
- Kanban (Hermes Kanban): A durable, SQLite-backed task management system for coordinating workflows across different agent profiles.
- Structured Handoff: The process of passing metadata, summaries, and context between tasks to ensure continuity without re-processing previous conversation logs.
- Dependency Promotion Engine: A mechanism that automatically moves tasks to "Ready" status once parent tasks are completed.
- Circuit Breaker & Crash Recovery: Resilience features that manage worker failures, prevent infinite retry loops, and recover tasks after process crashes.
- Dispatcher: The background component responsible for assigning tasks to appropriate agent profiles and managing the task lifecycle.
1. Hermes V0.11: The Interface Release
The V0.11 update focused on infrastructure, interface, and provider flexibility.
- CLI Overhaul: Rebuilt using React and Ink, featuring a "sticky composer," live streaming, a status bar, and improved visibility for sub-agents.
- Pluggable Transport Architecture: Decoupled provider communication into cleaner layers, enabling native AWS Bedrock support via the Converse API.
- Expanded Inference: Added support for NVIDIA NIM, RCAI, Google Gemini CLI, Vercel AI Gateway, and GPT 5.5 (via Codex OAUTH).
- Operational Tools: Introduced shell hooks, webhooks, direct delivery modes, and orchestrator-style sub-agents for smarter delegation.
2. Hermes V0.12: The Curator Release
The V0.12 update focused on autonomous maintenance and expanded ecosystem integration.
- Autonomous Curator: A background agent that grades, prunes, and consolidates the user's skill library.
- Self-Improvement Loop: Enhanced to be more rubric-based, allowing agents to update the specific skills they just utilized while inheriting parent runtime credentials.
- Platform Expansion: Added providers for Azure AI Foundry, Tencent TokenHub, and LM Studio.
- Gateway & Media: Introduced a pluggable gateway system (Microsoft Teams, Wechat), native Spotify tools, Google Meet plugins, and improved multimodal image routing.
- Performance: Achieved a ~57% improvement in "cold start" times.
3. The Kanban Framework: Coordination vs. Delegation
The video highlights a critical distinction in the Hermes architecture:
- Delegation: Best for short-lived, synchronous function calls where a parent agent waits for a sub-agent to return a result.
- Kanban: A durable work queue stored in
hermes/con.db. It is designed for long-running workflows that require persistence, human intervention, audit trails, and multi-role coordination.
Kanban Workflow States:
- Triage: Rough ideas needing refinement.
- Todo: Tasks waiting on dependencies or assignment.
- Ready: Assigned and waiting for the dispatcher.
- In Progress: Actively being executed by a worker.
- Blocked: Awaiting human input or triggered by a circuit breaker.
- Done: Completed work.
4. Key Use Cases for Kanban
- Solo Developer Feature Shipping: Uses the Dependency Promotion Engine. For example, an API implementation task remains in "Todo" until the "Design Schema" task is marked "Done," ensuring the worker has the necessary context.
- Fleet Farming: Managing independent tasks across specialist workers (e.g., translator, transcriber, copywriter). The "Lanes by Profile" view allows users to monitor parallel execution.
- Review Pipelines with Retries: Treats retry history as data. If a reviewer rejects a task, the engineer receives the feedback and the history of the previous attempt, allowing for iterative improvement rather than starting from scratch.
- Crash Recovery: If a worker process dies (e.g., OOM error), the dispatcher detects the failure, releases the claim, and moves the task back to "Ready" for a fresh attempt.
5. Technical Considerations & Limitations
- Scope: Kanban is designed as a single-host system. It uses a local SQLite database and is not intended to be a multi-server enterprise engine.
- Security: The dashboard should remain on
localhost. Exposing it to0.0.0.0makes plugin routes reachable from the network, posing a security risk. - Practicality: The system emphasizes that not every task needs Kanban. It is a tool for complexity, not a replacement for simple, one-off agent requests.
Synthesis
The evolution of Hermes from V0.11 to V0.12 marks a shift from "chat-based agent" to "persistent workflow system." By introducing the Kanban board, Hermes provides a structured way to manage complex, multi-step, and multi-role agentic processes. The core value proposition is the transition from ephemeral chat logs to durable, stateful task management where context, history, and failure recovery are treated as first-class data. This makes Hermes significantly more viable for real-world engineering and operational workflows.
Chat with this Video
AI-PoweredLoad the transcript when you're ready to chat so the initial page stays lighter.