Runtime Interface Adaptation for User Identity: Hermes Agent Onboarding

[djairjr]

Title: Runtime Interface Adaptation for User Identity: Hermes Agent Onboarding

Body:

Dear authors,

Your paper argues that adapting the runtime harness can improve frozen LLM agents more efficiently than updating model weights. I have been actively developing a complementary implementation that applies this principle to user identity and domain persistence.

My work: Hermes Agent Onboarding (github.com/djairjr/hermes-agent-onboarding) — a 6-stage meta-skill that builds persistent agent identity (local pgvector + Supabase) including fault logging, capability tracking, and generative domain ontology.

Alignment with your framework:

• identity_faults table implements your failure diagnosis concept (Section 4.1)
• Dynamic table/MCP generation per user is a form of your Environment Contract Layer (Section 4.3.1)
• Dual-tier storage (agent self-knowledge vs. user knowledge) extends harness adaptation to personalized, open-ended domains

I am currently extending this with a curator (pre-execution blocking) and async consolidation worker. The code is open source.

This may serve as a practical test case for your thesis outside deterministic benchmarks.

Best,
Djair Guilherme
https://github.com/djairjr/hermes-agent-onboarding

[Tianshi-Xu]

Thanks for sharing this! Hermes Agent Onboarding looks like an interesting and complementary direction, especially in extending runtime/interface adaptation toward user identity, persistence, and personalized agent memory.

One distinction is that LIFE-Harness currently focuses on deterministic, rule-governed environments where failures and improvements can be evaluated on held-out tasks with fixed success criteria. Your setting seems more open-ended and personalized, which is a valuable but also more challenging extension.

We appreciate the pointer and will take a closer look. This is an interesting direction for exploring harness adaptation beyond deterministic benchmarks.