Hermes Agent, the open-source agent framework from Nous Research, has overtaken OpenClaw on OpenRouter’s daily token consumption leaderboard, according to The Batch. The milestone arrives as OpenClaw crossed 145,000 GitHub stars, a pace that analyst Miles K. called “faster than any agent framework I’ve tracked this year.” Hermes Agent sits at 44,100 stars with 5,700 forks.
The token volume metric does not mean Hermes Agent has more users. Some Reddit users have noted that Hermes Agent consumes more tokens per task. But The Batch’s coverage highlights a more significant distinction: Hermes Agent’s ability to define and sharpen new skills automatically, positioning self-improvement as a core agentic capability rather than a feature checkbox.
Two Architectures, Two Bets
The competition is architectural, not feature-level. According to Miles K.’s analysis, the frameworks diverged at the foundation.
OpenClaw operates as a control-plane. A single long-lived Gateway daemon owns all messaging surfaces (WhatsApp, Telegram, Slack, Teams, Gmail). The LLM is a module plugged into routing infrastructure, constrained by design: system prompts under 1,000 tokens, four foundational tools (read, write, edit, bash), and an edit tool that rejects fuzzy matching entirely. Agent identity lives in Markdown files (AGENTS.md, SOUL.md, TOOLS.md) that a compliance officer can read, edit, and commit in Git. As Miles K. put it: “Control, predictability, auditability.”
Hermes Agent operates as an execution loop built around “do, learn, improve.” Its GEPA (Genetic-Pareto Prompt Evolution) system mutates the agent’s own prompts and skill descriptions against benchmarks like HotpotQA and IFBench via DSPy API calls. According to Miles K., agents using GEPA required up to 35 times fewer executions to solve complex problems. A “Dream Consolidation” background process reviews session transcripts after successful tasks and converts logical steps into reusable skills stored in procedural memory.
The Inspectability Tradeoff
The practical consequence, according to Miles K., is direct: “You are trading auditability for adaptability.” Hermes Agent’s capability matrix is distributed across SQLite FTS5 indexes, cached prompts, and accumulated GEPA outputs. If a compliance team asks what the agent knows how to do, “the full picture is not trivially inspectable.”
OpenClaw’s capability surface, by contrast, is readable text files. The community has built thousands of pre-vetted skills on ClawHub covering enterprise use cases from API documentation generation (Scribe) to pull request security scanning (Lens).
Infrastructure Economics
One dimension most comparisons miss, according to Miles K.: infrastructure cost at scale. OpenClaw requires a persistent Gateway host maintaining continuous WebSocket connections and local storage. Hermes Agent was designed for serverless persistence, with state that hibernates entirely when idle and instantiates on demand via platforms like Modal or Daytona. Idle compute costs drop to near zero.
For teams thinking about fleets of specialized agents running intermittently, that gap becomes significant.
The Self-Improvement Frontier
The Batch frames the broader shift: general-purpose agents are extending from stateless assistants toward systems that “accumulate experience, adapt to users, and automate ongoing work beyond isolated tasks.” Hermes Agent’s integration with Atropos, the reinforcement learning environment framework from Nous Research, positions it not just as a productivity tool but as training infrastructure for future model improvement.
The question for builders is whether their deployment context values the compliance-ready inspectability that OpenClaw provides or the compounding adaptability that Hermes Agent offers. Both frameworks support the same LLMs, the same SKILL.md format, and the same messaging integrations. The divergence is in what happens between tasks.