Building agents without harness engineering
5 hours ago
- #AI Agents
- #Hermes
- #API
- The author initially built an agent for prismvideos.com using Vercel AI SDK but faced competition from Higgsfield's Supercomputer agent built on Hermes.
- Realizing Hermes offered many built-in features for free, such as session management, tools, and automations, the author switched to using Hermes as a primitive.
- The new approach involved launching EC2 instances with Hono servers to create isolated Hermes agents in Docker containers per customer.
- This shift allowed the team to focus on domain-specific engineering instead of building agent infrastructure.
- The author argues that AI agent startups should differentiate by integrating with proprietary data and learning customer preferences, rather than building harnesses.
- An API is proposed to programmatically deploy Hermes agents with system prompts, tools, skills, and connectors in a single call.
- Comparisons are made with LangChain Managed Deep Agents and Claude Managed Agents, highlighting unique capabilities like automations and the Ralph Wiggum loop.