Agent systems / Featured system

Hermes / OpenClaw / Legion

The operational layer around agents: brokers, tools, browser sessions, collaboration protocols, local GPU services, and recovery.

Case study / Hermes / OpenClaw / Legion

01 / Context

Context

Agent runtimes become useful through an operational layer: credentials, queues, tools, browser sessions, local compute, channel health, recoverable state, collaboration conventions, and clear ownership of failures.

02 / Problem

Problem

The weakest boundary determines reliability. A browser session can expire, a provider can return the wrong file type, a GPU workload can contend for memory, a persona can drift, and a remote surface can be healthy but unreachable. Each failure needs a visible, recoverable path.

03 / System

System

The work spans a Hermes-backed media broker and MCP servers, structured vision loops, Xorvio management surfaces, genuine OpenClaw coordinator operations, a bidirectional Mac and agent protocol, authenticated browser workers, and a GPU-aware Legion control plane.

04 / Architecture

See the connective layer.

Architecture mode

System architecture / Hermes / OpenClaw / Legion

Agents submit typed work to broker and MCP boundaries; those boundaries schedule provider, browser, or local GPU execution; durable artifacts and redacted diagnostics return through explicit lifecycle states.

  1. Agent profile, schedule, or product request
  2. Tool, MCP, or broker contract
  3. Queue, concurrency, and credential boundary
  4. Provider, browser session, or local GPU service
  5. Native-file validation and durable asset
  6. Delivery, health evidence, and recovery state

05 / Hard parts

Where the work actually was

  1. 01

    Normalizing asynchronous provider, local GPU, and authenticated browser execution without leaking credential or topology details.

  2. 02

    Managing GPU contention and service lifecycle with desired-state behavior rather than brittle button-driven process control.

  3. 03

    Recovering OpenClaw persona and memory state from evidence while preserving healthy live configuration.

  4. 04

    Separating original integration work from upstream core authorship and keeping experimental work visibly experimental.

06 / What changed or was learned

The field note

Agent quality depends on the systems around the model. Durable assets, observable queues, explicit credentials, healthy browser sessions, resource-aware scheduling, and snapshot-based recovery turn impressive demos into operable tools.

07 / Evidence

Proof

Safe public evidence, translated from tests and runtime validation rather than raw internal logs.

TypeEvidenceWhat it establishes
runtimeAsynchronous media broker

Normalized hosted APIs, local GPU execution, editing, video, and authenticated browser workers through queued jobs and durable asset delivery.

operationsOpenClaw recovery

Restored coordinator persona and memory behavior by comparing live configuration with verified snapshots before applying focused changes.

runtimeGPU workload control

Used desired-state reconciliation and GPU-aware mutual exclusion across chat, image, music, audio, embedding, and specialty workloads.

artifactAgent collaboration

Built bidirectional handoff conventions with indexes, markers, synchronized files, service watchers, and agent-created visual artifacts.

08 / Stack

Technology in service of the system

  • Hermes
  • OpenClaw
  • MCP
  • React
  • Node.js
  • FastAPI
  • Python
  • Docker Compose
  • systemd
  • launchd
  • SSH
  • rsync
  • CDP
  • NVIDIA CUDA
  • PyTorch
  • FlashAttention
  • Triton

10 / Status and boundaries

Built with the edges visible.