Finding
Broad web discovery can turn into an unbounded research spiral unless Hermes defines the question, source categories, stopping criteria, and preservation path before searching.
Current
A typical Hermes installation can search and extract enough web content to discover useful unknowns, but the weak point is scope control. Without a research frame, the agent may collect many interesting pages, mix official sources with commentary, and leave the final recommendation hard to verify or reuse. The achievement signals that the installation is capable of deep discovery; the maturity step is making that discovery bounded, source-ranked, and decision-ready.
Suggested
- Add a bounded discovery template for research tasks. Exact change: create or update a
research-workflowskill section named “Rabbit Hole Guardrails” with required fields for research question, allowed source categories, excluded categories, maximum search rounds, minimum source count, and stop condition. - Add source category locking to research prompts. Exact change: patch the recurring research cron prompt or runbook to require categories such as “official docs,” “release notes,” “GitHub issues,” “community examples,” and “vendor pricing,” then instruct the agent to search within those categories instead of following every related lead.
- Add a preservation and verification habit before recommendations. Exact change: update the research wiki runbook at
/opt/data/home/research/README.mdto require asource-map.md,findings.md, anddecision.mdfor every broad discovery task, with each major claim linked to a URL or saved file path before the final recommendation is written.
Impact
This turns web research from curiosity-driven browsing into a controlled discovery workflow. Hermes still finds unknown opportunities, but it avoids drowning the operator in loosely related pages. The result is easier to audit, easier to resume later, and safer to convert into cron monitors, skills, or implementation plans.
Effort
Medium — it requires a small skill/runbook update plus discipline in research prompts, but no new infrastructure.
Public page note
Safe public content includes the achievement meaning, research guardrails, generic source categories, and maturity recommendation. Internal-only content includes private research topics, raw chat excerpts, unpublished source maps, sensitive URLs, credentials, logs, customer data, and any live filesystem paths that reveal private operational details.