This article details the process of upgrading a specialist agent within the Hermes AI framework, focusing on enhancing the "researcher" agent's capabilities. The upgrade involves integrating new skills, connecting to external tools like Zotero and Consensus, and demonstrating how these improvements boost the efficiency and quality of an agent team. The ultimate goal is to create more robust and reliable AI workflows for knowledge management and research.

Enhancing the Researcher Agent

The researcher agent, a foundational member of the AI team, was upgraded to improve its ability to conduct in-depth research and collaborate with other agents. This upgrade focused on three key areas:

1. LLM Wiki Research Skill: This skill introduces different research "lanes" or "lenses," allowing the researcher to focus on academic sources, current reality, or a combination of both. It also ensures that research findings are grounded with citations and presented in a readable report, which is then handed off to the orchestrator agent. This skill leverages the built-in "grounded citation" skill for footnote and citation tracking.
2. Zotero Connection: This integration grants the researcher agent read-only access to a personal Zotero library, a reference management system used for collecting, organizing, and citing research sources. The researcher can search and inspect existing local library items. This connection utilizes deterministic tooling, validating information from Zotero and saving a copy in a research workspace connected to the LLM wiki. To enable this, the Zotero setting "allow other applications on this computer to communicate with Zotero" must be enabled.
3. Consensus Academic Discovery (MCP): This skill enables the researcher to perform bounded academic paper discovery through Consensus, an external platform specializing in academic literature. The skill preserves useful paper metadata, records search limits, and requires following original links for inspection before making claims. It does not modify Zotero or the LLM wiki. The connection is established through the "MCP" (likely referring to a specific integration protocol or service) capability within Hermes, authenticated via the Consensus platform.

Workflow Integration and Testing

The upgraded researcher agent was tested within a "skills-based research team" group chat, comprising the researcher, an orchestrator, and a librarian agent. The process involved:

• Task Delegation: The orchestrator agent was tasked with delegating a request to the researcher for a combined report on "agentic AI safety." This report was to compare and contrast academic research from Consensus with existing knowledge in Zotero.
• Research Execution: The researcher agent executed the task, querying both Zotero and Consensus. It identified relevant papers from Zotero, including sources on "memory poisoning," and discovered academic papers through Consensus.
• Quality Assurance: The researcher presented a markdown report to the orchestrator. The orchestrator performed a quality assurance check, requesting minor provenance corrections for details from Consensus and Zotero.
• Librarian Ingestion: Upon approval, the orchestrator passed the report to the librarian agent. The librarian reviewed the report, proposed its inclusion in the LLM wiki, and connected the new concept of "agentic memory poisoning" to existing entries like "agentic memory" and "multi-agent workflows." The librarian's process includes a "review gated ingest" and a "human in the loop" review skill.

Outcomes and Future Directions

The upgraded researcher agent demonstrated a significant improvement in the quality and depth of research. The integration with Zotero and Consensus provided access to a broader and more credible set of sources. The collaborative workflow between the researcher, orchestrator, and librarian resulted in a well-grounded report that was successfully ingested into the LLM wiki.

Future enhancements planned include:

• Automation: Implementing a "goal" feature to allow the researcher to run on a loop or schedule weekly updates.
• Efficiency: Improving the efficiency and consistency of handoffs and checks between agents.
• Scalability: Utilizing the Hermes Kanban feature and upgrading the orchestrator agent with more tools to create a reproducible workflow for scaling agent interactions.