0005, Defer subagent fan-out pending a design pass¶
- Status: Deferred (2026-06-02)
- Deciders: Ariadne maintainers
Context¶
The best-practice research names an orchestrator-worker design, a lead agent dispatching parallel, context-isolated subagents, one per source, as the headline architecture. The obvious next step would be to fan out the graph and SQL retrieval into separate subagents.
Decision drivers¶
- Cross-store reconciliation (corroborate facts, flag conflicts about the same entity) is a shared-context task.
- Provenance is the spine: a parent-side
PostToolUsehook assigns each evidence call agNid that citations resolve to. - Add cost and complexity only when they buy something the current slice needs.
Considered options¶
- Fan out now: one subagent per store, run in parallel.
- Defer: keep the single lead agent with both connectors until a design resolves the conflicts below.
Decision¶
Defer. Naive fan-out collides with two load-bearing properties of the current slice:
- Shared context. Reconciliation needs both stores' evidence in one context; the research explicitly flags multi-agent fan-out as not generalising to shared-context tasks.
- Provenance. Subagents run in isolated contexts and return only a summary,
their raw tool calls never reach the parent's
gNhook, so citations would attach to a worker's prose instead of to evidence entries.
A correct design (workers retrieve in parallel and return pre-cited evidence; the lead reconciles) is a real redesign of the provenance layer. It is not blocked on research, it needs a focused design pass.
Consequences¶
- The 2-store slice keeps the single-agent design and already scores
grounded=Trueon the eval harness. - Revisit when store count or context pressure justifies the cost (multi-agent fan-out runs ~15× the tokens of a single agent).
- When revisited, the provenance redesign is the gating work, not the parallelism itself.