Hermes DevLoop
Human-Mentored Agentic Engineering
The Project
Hermes DevLoop began with a question: Can an AI coding agent develop better engineering judgment through mentorship?
Coding agents are increasingly capable of implementing software, but producing code is only part of engineering. Understanding requirements, recognizing risks, evaluating alternatives, maintaining systems, and knowing when to ask for help are equally important.
I'm exploring whether an agent can improve these capabilities through sustained interaction with an experienced human engineer.
The objective isn't to eliminate human involvement. It's to understand how agents and engineers can work together, how responsibility might evolve, and whether improvements in an agent's decision-making can be demonstrated.
Building the Environment
The project uses Hermes as the foundation for a supervised software development environment.
The agent works within an isolated, persistent workspace where it can investigate engineering problems, examine source code, develop solutions, perform available verification, and propose changes through GitHub.
The environment is designed to support increasingly capable agent participation without giving the agent unrestricted control.
Human engineers retain responsibility for architectural direction, mentorship, consequential decisions, code acceptance, and production changes.
Supporting capabilities include development isolation, source control integration, controlled permissions, automated verification, review workflows, and observability.
These capabilities are being developed incrementally as the experiment progresses.
The Mentorship Experiment
The longer-term objective is to explore whether sustained mentorship changes how an agent approaches engineering problems.
A useful engineering assistant should do more than follow a prescribed sequence of steps. It should recognize when information is incomplete, identify potential consequences, consider alternatives, and respond appropriately to uncertainty.
I'm interested in whether feedback on these decisions can influence future behavior, including when the agent encounters problems it hasn't previously seen.
An important part of the experiment is distinguishing genuine, transferable improvement from task-specific instruction following, retained memory, or additional automation.
Engineering Principles
The environment is built around several principles:
- Autonomy with boundaries. Agents should have sufficient freedom to investigate and solve problems without unrestricted access to systems or credentials.
- Human accountability. Architectural direction, mentorship, review, and consequential decisions remain human responsibilities.
- Evidence over assumptions. Tests, operational evidence, and independent review should support claims about system behavior.
- Learning through feedback. Engineering reviews should provide opportunities to improve future decisions, not simply accept or reject code.
- Measured progress. Changes in correctness, judgment, cost, speed, and human intervention should be evaluated rather than assumed.
These principles guide the experiment without predetermining its outcome.
An Open Engineering Experiment
Hermes DevLoop is being developed publicly.
The repository includes source code, technical documentation, engineering issues, and pull requests. These provide visibility into the implementation, design decisions, discussions, corrections, and ongoing work.
The project remains experimental. The environment supports supervised engineering activities, while the broader questions about mentorship, transferable judgment, and increasing agent responsibility remain under investigation.