Deployment of an autonomous background AI agent capable of orchestrating multi-step engineering tasks, environment self-healing, and persistent context management through SQLite.
The Challenge: Reliable, Always-On Agent Execution Without Vendor Lock-in
Standard cloud AI platforms charge steep premiums for long-running workflows and restrict access to low-level shell commands and local file systems. The goal was to build a dedicated, self-hosted execution environment running continuously on an Oracle Cloud instance, governed by Traefik reverse-proxy and orchestrated with Dokploy.
Target Architecture & Tool Registry
Hermes operates with an asynchronous execution loop connected to an isolated Docker engine. When a task is accepted, subagents are spawned with scoped tool permissions: Docker container control, shell commands, file read/write, and Dokploy API endpoints. All state mutations and execution trees are committed to an encrypted SQLite database.
Key Engineering Challenges
- 1
Preventing infinite reasoning loops by implementing budget-aware halting heuristics and tool call throttles.
- 2
Preserving session memory across daemon reboots without bloating prompt context window.
- 3
Hardening reverse-proxy ingress with Cloudflare mTLS and Dokploy token authentication.
Operational Metrics
- 99.98%
- Uptime
- 1,800+
- Daily tool calls
- < 4ms
- Context retrieval latency
- 4 concurrent
- Active subagents