Hermes, an architecture pilot for AI agents
A secure-architecture pilot for running autonomous AI agents in isolated containers, with a budget and checks on every run.
- Case study
- 07 / 07
- Projects
- AI agents · isolation · controls
- Type
- Personal project
- Role
- Individual work; detailed design, code and tests largely produced with Claude Code
- Period
- Work from 3 to 15.11.2026
- Status
- Experimental (pilot)
01
Overview
Hermes is a pilot, not a running platform: a test bed for having autonomous AI agents carry out tasks within tight limits. Each run happens in a throwaway container, one after another, on isolated Docker networks, behind a proxy that only allows the expected domains, with a budget cap.
02
Context
Work carried out from 3 to 15 November 2026. The aim was not to automate a job, but to check whether tasks can be handed to autonomous agents with guardrails that can be verified.
03
Challenge
Autonomous agents are easy to demonstrate and hard to operate. The difficulty lies in the frame around each run: choosing the right agent for the task, isolating it, limiting its network access and its budget, handling its failures, and knowing when a human needs to step in.
04
Objectives
- Isolate every run
- Restrict the network to allowed domains
- Cap the cost of each run
- Keep a record of what each agent does
05
Technical decisions
Sequential runs in throwaway containers; isolated Docker networks; a Squid proxy allowing only two domains; a hardened launcher; a budget cap per run; daily encrypted backups.
06
Architecture
Launcher → throwaway container (one agent, one run) → isolated network → Squid proxy (domain allowlist) → model.
07
Technologies
- Linux
- Docker / Docker Compose
- Squid
- Python
- Shell
08
Implementation
13 runs and 44 commits from 3 to 15 November 2026. Each of the 7 registered agents ran at least once, four of them only once.
09
Results
The controls stopped 3 runs: a size limit, a budget refusal and a refused address extraction. These were not attacks, but cases the guardrails were meant to stop. Estimated model cost: USD 4.20 in total (an estimate, not reconciled with an invoice).
10
Limits
This is a pilot. Some network protections were tested on a bench but not yet re-checked in real conditions after the last change; the container starts with elevated privileges before handing over to an unprivileged user; a full restore has not been proven.
11
My role
I set the goals and constraints, made the decisions, directed Claude Code, approved changes and results, and drove the step-by-step improvement. The detailed design and much of the code and tests were produced with Claude Code, under my direction.
12
Lessons learned
An autonomous agent becomes workable when its execution is controlled and a human can step in.