subagent-research-reliabilitylisted

# Subagent Research Reliability ## Problem Dispatching subagents for web research has two silent failure modes that both produce confident-looking but worthless output: 1. **The agent can't actually search.** Its `tools:` grant lacks WebSearch/WebFetch even though its description says "performs web research." It returns a no-op or fabricates citations from training data. 2. **The agent searches but hallucinates sources.** LLM research output contains plausible-but-fake citations, CVE IDs, arXiv IDs, dates, or fabricated cross-references (e.g., a real CVE attributed to a paper that never cites it). Both are invisible if you trust the returned text. Both were hit and caught in one session (2026-05-28 — see [EVIDENCE.md](EVIDENCE.md)). ## Context / Trigger Conditions - About to call the Agent tool with a custom `subagent_type` for web research / literature / threat-intel / market scan. - The agent's description claims a web capability. - Curating findings (citations, CVE/arXiv IDs, advisory URLs, dates) returned by a research subagent before acting on them. ## Solution ### Check 1 — Pre-dispatch: verify the agent's tool grant matches the task Before dispatching `subagent_type=X` for web research, confirm X actually has the web tools: - Read `.claude/agents/X.md` (or `~/.claude/agents/X.md`) frontmatter `tools:`. If it lists only `Read, Bash, Grep` (no `WebSearch`/`WebFetch`), the agent **cannot web-search** regardless of its description. - Observed: