Hermes Agent & OpenClaw

Hermes Agent & OpenClaw: How We Build Agentic AI at ESS ENN

The best way to learn how to build dependable AI agents is to build them — repeatedly, for real workloads, and to feel every sharp edge yourself. That is the thinking behind two of our in-house agentic AI initiatives, Hermes Agent and OpenClaw. They are not products we sell off the shelf; they are how we pressure-test the patterns, guardrails, and architecture that go into the custom agents we build for clients through our AI agent development practice. This post opens the hood on both, and on the principles they have taught us.

New to the topic? Our primer What Is Agentic AI? covers the fundamentals, and our production playbook covers the engineering. Here we focus on what building our own agents has taught us.

Hermes Agent: Coordination and Communication

Named for the messenger of the gods, Hermes Agent is our initiative focused on the communication- and coordination-heavy work that fills so many business days: reading incoming messages and requests, understanding intent, drafting and routing responses, scheduling and following up, and stitching together the several tools a single task touches. It is the kind of workflow where a chatbot is too passive and a rigid script is too brittle — exactly the gap an agent fills.

The hardest lessons from Hermes were not about intelligence but about trust and control. An agent that drafts a customer reply is helpful; an agent that sends one unsupervised is a liability until it has earned that right. So Hermes is built around human-in-the-loop checkpoints: it does the laborious legwork — gather context, draft, propose next steps — and a person approves anything that leaves the building. That single design choice, autonomy for the tedious parts and human judgement for the consequential parts, is now a default in every client agent we ship.

OpenClaw: Giving Agents Safe Hands

If Hermes is about what an agent says, OpenClaw is about what an agent does. It is our exploration of tool use — the machinery that lets an agent reach out and act on the world through external tools, APIs, and data sources. An agent without tools is just a clever conversationalist; an agent with tools can query systems, fetch data, trigger workflows, and complete real tasks end to end. OpenClaw is where we work out how to grant those capabilities safely.

The recurring theme is least privilege and observability. Every tool an agent can reach is a door, and every door is a risk if left unguarded. OpenClaw taught us to scope each tool to the minimum permission it needs, to put approval gates in front of anything irreversible or costly, to sanitise inputs against prompt-injection attempts that try to hijack the agent through poisoned data, and to log every single action with its reasoning so nothing the agent does is a mystery after the fact. These patterns map directly onto standards like the Model Context Protocol (MCP) that we use in production builds.

The Principles These Initiatives Crystallised

Across both efforts, a consistent set of design principles emerged — the same ones we now apply to every custom agent:

• Production-first, not demo-first. We design for the thousandth run, not the first — with fallbacks, retries, and graceful degradation.
• Autonomy where it's safe, humans where it matters. Routine work runs unattended; consequential actions wait for approval.
• Least-privilege tools. Every capability is scoped to the minimum, and risky actions are gated and logged.
• Model-agnostic engineering. We build to swap between frontier and open models as the field moves, avoiding lock-in.
• Measure everything. Evaluation suites and observability turn improvement into evidence instead of guesswork.

From Our Lab to Your Workflow

The point of Hermes Agent and OpenClaw is not to hand you a generic bot — it is to make sure that when we build your agent, it is grounded in patterns we have already stress-tested. Some clients want a focused single-purpose agent; others need an orchestrated team of agents; many want to automate an existing business process end to end, which leads naturally into agentic process automation. In every case, the agent is usually wrapped in a custom interface, APIs, and data pipelines built by our AI applications team, so it slots cleanly into how your business actually works.

Frequently Asked Questions

Related reading: What Is Agentic AI? and our playbook for deploying custom AI agents to production.

At ESS ENN Associates, the patterns behind Hermes Agent and OpenClaw power the production agents we build through our AI agent development and agentic process automation services. To put a custom agent to work in your business — contact us for a free consultation.