Use case: multi-agent deal intelligence vault (150+ files, 5 autonomous agents)

[bizbuilderai]

We run a structured markdown vault (Obsidian-style, ~150 files) for a B2B SaaS project — company dossiers, people profiles, outreach logs, signal processing, daily journals. Multiple autonomous Claude Code agents write/read files on a schedule.

Tried knowledge-graph on it this week. Some observations and questions from real usage:

What works well:

• Path finding between company → person → outreach → signal nodes maps exactly to what we call "resonance checks" (tracing how a market signal connects to a product capability through the relationship chain). This is the killer feature for us.
• Community detection surfaced clusters we hadn't named — targets naturally grouped by industry vertical and relationship type. Useful for finding shared angles across targets.
• Incremental indexing matters a lot when agents are writing files throughout the day.

Where we hit edges:

1. YAML frontmatter depth. Our files have structured metadata in frontmatter (classifications, multi-value fields, nested properties). The parser extracts frontmatter via gray-matter, but rich metadata doesn't seem to flow into edge weights or node properties. Surfacing typed frontmatter fields in kg_node output would make the graph much richer than untyped connections.

2. Temporal edge weighting. A wiki link from yesterday's journal entry is more "alive" than one from three weeks ago. For deal intelligence, recency matters. Is there a path toward time-weighted queries, or is that out of scope?

3. Directional reasoning chains. Our vault has a specific traversal pattern across node types (e.g. signal → person → company → capability → proof point). The graph treats all links as undirected. Typed or directional edges would let agents run structured queries like "given this signal, what's the shortest path to a proof point?"

4. Bridge nodes = networking gold. Betweenness centrality on our people graph identified the exact same "key connectors" we'd found manually through weeks of outreach. Strong validation of the algorithm for this use case.

Not filing these as feature requests — more sharing what a production vault looks like when agents are the primary readers. Happy to share more details on vault structure patterns if useful for testing edge cases.