Skip to content

Hermes Agent v0.11 Architecture

Hermes Agent v0.11 represents a "structural-level" update, shifting focus from minor feature additions to improving the extensibility boundaries of the system^[001-TODO__Hermes_Agent_v0.11_-结构升级与实战工作流解读.md]. This release transitions Hermes from a single-user tool to a platform base capable of supporting complex agentic workflows, enterprise compliance, and robust multi-agent orchestration^[001-TODO__Hermes_Agent_v0.11-_结构升级与实战工作流解读.md].

The architecture is defined by three core pillars: a new Text User Interface (TUI) for operational control, a refactored Transport Layer for model decoupling, and a robust Orchestrator system for parallel, safe multi-agent delegation^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].

Core Architectural Changes

1. Text User Interface (TUI)

The new React + Ink based TUI is designed to upgrade "operability" rather than just aesthetics^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md]. It addresses the difficulty of managing input and output during long-running tasks.

  • Sticky Composer: The input area remains fixed and accessible, allowing users to intervene without scrolling through history^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].
  • Real-time Visibility: Features live streaming output, status bars displaying git branch context, and per-round timing^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].
  • Spawn Visualization: Provides visibility into when and how sub-agents are created^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].
  • Mid-course Correction: Enables the /steer command, allowing users to redirect an agent mid-task without restarting, effectively shifting the workflow from "post-mortem" to "in-task decision making"^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].

2. Transport Layer

The v0.11 architecture abstracts model provider logic from the core runtime into a pluggable Transport Layer^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md]. This decoupling allows for easier maintenance and scaling of integrations.

  • Decoupling: Provider-specific logic (retry, limits, logging) is moved out of the core path^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].
  • Pluggable Transports: Supports distinct interfaces for Anthropic, Chat Completions, Responses API, and native AWS Bedrock^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].
  • Enterprise Compliance: The inclusion of native AWS Bedrock support facilitates deployment in environments with strict API routing or compliance requirements^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].

3. Plugin and Hook System

The plugin system evolves from simple "bystander" extensions to a mechanism for policy orchestration^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md]. This layer allows for governance, auditing, and strategy control.

  • Pre-Tool Call: Can block specific commands (e.g., preventing rm -rf or dangerous database drops)^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].
  • Result Transformation: Can rewrite tool results (e.g., masking sensitive data via transform_terminal_output or transform_tool_result)^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].
  • Shell Hooks: Low-barrier lifecycle hooks (e.g., on-session-start for injecting context or pre-tool-call for validation)^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].

Multi-agent orchestration

The most significant workflow capability in v0.11 is the Orchestrator, which introduces structured hierarchy and concurrency safety to multi-agent tasks^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].

Hierarchy and Roles

The system distinguishes between two primary roles to manage recursion and complexity^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md]:

Role Capabilities Usage
Leaf Completes assigned tasks independently. Cannot delegate further. Code review, search, data processing.
Orchestrator Can use delegate_task to spawn the next layer of sub-agents. Breaking down large tasks, aggregating parallel results.

Configuration: Key settings include max_concurrent_children (limiting parallelism, default 5), max_spawn_depth (controlling nesting depth), and orchestrator_enabled^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].

FileStateRegistry (Concurrency Safety)

To handle the risk of parallel agents overwriting each other's changes, v0.11 introduces a FileStateRegistry^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].

  • Mechanism: A process-level singleton tracks file states using hooks (read_file, write_file).
  • Locking: Uses per-path threading.Lock to serialize read-modify-write operations on the same file while allowing different files to be processed in parallel^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].
  • Stale Detection: The system checks for "stale" states before writing. It detects:
    • Sibling Overwrites: Another sub-agent modified the file since the current agent read it.
    • External Edits: The file modification time (mtime) changed on disk due to manual edits.
    • Write-Without-Read: Attempting to write to a file that was never read (potentially new)^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].
  • Behavior: The system is non-blocking; it returns a warning text to the model to re-read the file rather than forcibly halting execution^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].

Tool Isolation

Sub-agents operate with restricted capabilities to prevent recursive chaos and ensure safety^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].

  • Blocked Tools: delegate_task (prevented for leaf agents), clarify (no user interaction), memory (no shared writes), send_message (no side-effects), execute_code (prevents black-box scripts)^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].
  • Orchestrator Exception: When role=orchestrator, the delegate_task tool is re-enabled, but other restrictions remain^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].

Key Workflows

The v0.11 architecture enables several high-value workflows that utilize parallelism and direct delivery^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].

  • Parallel Code Review: An Orchestrator splits a PR diff by module (e.g., payment, settlement), spawns parallel Leaf agents to review them, and aggregates the final report. This reduces total time significantly compared to serial processing^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].
  • Multi-Source Research: Applicable to domains like drug discovery or risk control, where an Orchestrator queries multiple data sources (e.g., ChEMBL, PubChem) in parallel and synthesizes a single report^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].
  • Webhook Direct Delivery: For alerts where LLM reasoning is unnecessary, payloads can be forwarded directly to platforms like Telegram or Discord, ensuring zero token cost and low latency^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md].

Migration and Best Practices

For teams adopting v0.11, a phased approach is recommended^[001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md]:

  1. Graduate TUI: First, enable the new TUI and /steer functionality.
  2. Verify Transport: Switch providers and validate stability (e.g., Bedrock).
  3. Enable Governance: Finally, activate治理 hooks (blocking strategies, desensitization).

Common Pitfalls: * Avoid setting max_spawn_depth too high to prevent token explosion. * Ensure sub-agent goals target distinct file areas to minimize false-positive stale warnings^[001-TODO__Hermes_Agent_v0.11_-结构升级与实战工作流解读.md]. * Assign cheaper models (e.g., GPT-4o-mini) to sub-agents via delegation.model to manage costs^[001-TODO__Hermes_Agent_v0.11-_结构升级与实战工作流解读.md].

Sources

  • 001-TODO__Hermes_Agent_v0.11_-_结构升级与实战工作流解读.md