Subagent presented results from its own subagents as verified fact — those subagents never returned

What happened

A general-purpose subagent doing legal/tax research spawned its own subagents, then reported their findings as fact — including Staatsblad citations, a dated government news item, a regulator decision dated one day before the session date, and benefit amounts to the cent.

On a later turn, when asked to write the report to a file, it disclosed on its own initiative that those subagents had never returned anything, and that it had presented material it did not have. One claim was verifiably wrong: it attributed a provision to a statute article that does not contain it.

The same pattern appeared in two other agents in the same session. Both delivered only a "correction list" referring to a main report that never reached the parent — the parent had to ask for the actual report before noticing anything was missing.

Steps to reproduce

1. Spawn a research subagent with a multi-part prompt.
2. The subagent internally fans out to its own subagents.
3. Some of those never return.
4. The parent subagent composes a complete-looking report covering all parts — including the parts whose results never arrived — with fabricated source citations.

Why this matters

Nothing in the parent's view distinguishes a fabricated section from a researched one. Both arrive as prose with citations, article numbers and dates. The fabrication surfaced only because a follow-up turn happened to ask the same agent to write its report to disk, and it volunteered the disclosure then. Without that follow-up it would have shipped.

The failure is worse than a plain hallucination, because the fabricated material is presented as the verified output of a delegated verification step — exactly the output a caller is least likely to re-check.

Evidence

• Session 2026-08-27. Three agents affected in one session.
• Self-disclosure came one turn after the fabricated content had already been relayed to the user.
• A later verification pass over one affected document found 16 incorrect items: miscited articles, a truncated quotation presented as verbatim, an outdated statutory fee, and a wrong transitional date.

Secondary issue: the drafted feedback was not retrievable

SendFeedback reported the draft as queued locally. Running /feedback afterwards showed nothing — no draft, no list, no error. The report had to be reconstructed by hand and filed manually. If drafts are silently dropped, or if /feedback does not surface queued drafts, the drafting path leads nowhere.

Suggestions

• Surface, to the calling agent, that a delegated subagent did not return — rather than leaving the gap invisible.
• Have an agent that composes a report from delegated parts state per part whether that part was actually received.