What HermisAgent is
HermisAgent is the runtime that executes long-running, multi-step AI agents inside Open Agent. It owns the plan, the memory, the tool catalogue, and the safety layer. Open Agent is the product surface. HermisAgent is the engine underneath.
Three problems we set out to solve
- Long-running work. Real tasks take minutes, hours, or days. The runtime must survive restarts, retries, and human pauses.
- Tool sprawl. Production agents need access to dozens of tools. The runtime must keep the catalogue searchable and let the model pick wisely.
- Trust. Operators need to understand what the agent did and why. Every decision must be auditable.
The five components
- Planner. Turns a goal into a sequence of steps. Re-plans when reality disagrees with the plan.
- Memory. Short-term scratchpad plus long-term episodic memory backed by a vector store.
- Tool catalogue. Tools described with structured schemas. Retrieved on demand so the model only sees what it needs.
- Executor. Runs tool calls with retries, rate limits, and per-step cost limits.
- Safety layer. Input filters, output filters, and human approval for irreversible actions.
Why we built our own runtime
We tried wiring agents on top of off-the-shelf frameworks. Every one broke at the seams when the workflow lasted longer than a single API call. HermisAgent was the smallest engine we could write that handled real production state.
What our customers ship on top of it
- Recruiting agents that screen candidates, schedule calls, and update the ATS.
- Sales research agents that prepare briefing notes before every customer call.
- Compliance agents that monitor incoming documents and flag policy drift.
What we tell new builders: Start with a tight goal. Pick three tools. Add an approval step before anything destructive. Ship that, measure, and only then expand the tool catalogue.
Related reading
For the lower-level pattern, see LangGraph for Stateful Agents. For the broader stack, see The LLM Stack in 2026.
