Subagent research reliability

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).

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: a research-scout agent described as "Performs web research ... with citations" had tools: Read, Bash, Grep. Dispatching it for a live web beat would no-op or fabricate.
• Fix: dispatch general-purpose (tools *, includes WebSearch/WebFetch) with the research protocol embedded in the prompt; OR add WebSearch/WebFetch to the agent's tools:. The description is not the capability — the frontmatter is.

Check 2 — Post-return: verify citations with a dedicated pass

Never treat research-subagent citations as confirmed. After research returns, dispatch a SEPARATE verification subagent (general-purpose) whose ONLY job is to WebFetch each source URL and report, per URL: VERIFIED (resolves + content matches) / PARTIAL (resolves but a detail differs) / UNRESOLVED (404) / UNCONFIRMABLE. Tell it explicitly: verify source existence + content match, do NOT assess the threat/quality.

• It catches dead URLs, fabricated CVE/arXiv IDs, and fabricated cross-references (a real ID bolted onto the wrong source).
• Observed catch: real Cursor CVEs (CVE-2025-54136/54135) attributed to an arXiv paper that doesn't cite them.
• Only findings that survive verification become actionable (eligible to become a tracked finding, a security-finding candidate, or a dependency action).

Verification

• Pre-dispatch: the dispatched agent type's tools: includes the needed web tool, OR you used general-purpose.
• Post-return: every actioned citation carries a VERIFIED/PARTIAL verdict; UNRESOLVED or fabricated ones are dropped or annotated in place.

Example

Threat-watch beat: 3 general-purpose subagents (NOT the web-toolless research-scout agent) ran clusters in parallel → 9 findings. A 4th general-purpose subagent WebFetched all 9 source URLs → 7 VERIFIED, 2 PARTIAL, and flagged one fabricated CVE-to-paper linkage. Only verified findings were recorded as actionable; the fabrication was corrected in the log.

Notes

• Claude Code discipline (custom agents, subagent_type, WebSearch/WebFetch tools, agent frontmatter).
• A verification subagent is cheap relative to the cost of acting on a hallucinated citation (e.g., filing a security finding against a fake CVE, or bumping a dependency for a CVE that doesn't exist).
• Background dispatch + notify-on-complete pairs well (don't block; curate on return) — but echo the verdict inline, never leave it only in a background temp file.
• A project's research-agent SKILL may define the finding schema, but the AGENT definition is where the tool-capability trap lives — they can drift apart. See also superpowers:subagent-driven-development (spec/review loops) for the broader dispatch discipline family.