The AI Agent Landscape — Hermes Agent, OpenClaw, and Kilocode

AI agents are tools that go beyond simple question-and-answer interactions. They can use tools, take multi-step actions, read and write files, run code, and interact with external services. The ecosystem is evolving fast, and the terminology is loose. This post maps out where a few specific tools fit and how they differ.

What makes something an "agent"

Before looking at specific tools, it's worth being clear about the term. In practical terms, an AI agent is a system where a model can take actions — not just generate text — and use the results of those actions to inform what it does next. A simple example: an agent that can read a file, summarise it, search the web for related information, and write a report, without you intervening between each step.

The key ingredients are: a capable model, a tool use / function calling interface, and an orchestration layer that manages the loop between model output and tool execution.

Hermes Agent

Hermes refers to the NousResearch Hermes model family — specifically models fine-tuned for strong tool use, structured output, and agentic workflows. Hermes models are available on Hugging Face and run well locally via Ollama or llama.cpp.

What makes Hermes notable for local agent work is its reliable function-calling capability. Many open-weight models support tool use in theory but produce malformed JSON or hallucinate tool calls in practice. Hermes models are specifically tuned to avoid this, making them a practical choice as the reasoning core of a locally-running agent.

If you're building a custom agent framework or need a local model that consistently follows tool schemas, Hermes is worth evaluating over general-purpose alternatives.

Kilocode

Kilocode is a VS Code extension that brings AI coding assistance into your editor with support for local models via Ollama. It's positioned as an open-source alternative to GitHub Copilot or Cursor — one that you can point at a model you control rather than a hosted API.

Where Kilocode differentiates itself is in its agentic capabilities within the editor: it can read your project structure, understand file context beyond the current file, execute terminal commands, and make multi-file edits. It supports the Model Context Protocol (MCP), which means it can be extended with custom tools.

For developers who want the Cursor-style experience with a local model, Kilocode is one of the more capable options available today. The trade-off is that the quality of suggestions depends heavily on the model you're running locally — a 7B model won't match a frontier model for complex reasoning tasks, but for code completion and refactoring it performs well.

OpenClaw

OpenClaw is a newer entrant focused on local AI agent orchestration — building pipelines where multiple models or tools work together to complete complex tasks. It provides a framework for defining agent workflows, managing context between steps, and integrating with local model servers.

It's more of a developer tool than an end-user product: you define what the agent can do, what models it uses, and how it handles failures. The value is in the orchestration layer rather than a polished interface.

If you're building something beyond a single model interaction — multi-step research pipelines, automated code review workflows, document processing chains — OpenClaw is worth exploring as an alternative to building the orchestration logic yourself.

How they fit together

These tools aren't competitors in the traditional sense — they operate at different layers of the stack. Hermes is a model choice. Kilocode is an editor integration. OpenClaw is an orchestration framework. A reasonable local agent setup might use a Hermes model running via Ollama, accessed by Kilocode for in-editor assistance, with OpenClaw handling automated background workflows.

The broader takeaway: the local AI agent space is fragmented and moving fast. It's worth experimenting with specific tools for your actual use cases rather than trying to settle on a definitive stack. The final post in this series ties everything together into one coherent setup.