Deep Research

deep-research skill — Usefulness & Uniqueness

Sep 8, 2026 · 70+ sources · Score: 12/12

Summary

The deep-research skill is genuinely useful for developers who need auditable, file-based research artifacts that live alongside code. Its uniqueness is integrative, not algorithmic: no competitor ships the full five-pattern stack (parallel DAG subagents, per-claim atomic evidence, plan pre-mortem, self-scoring loop, IDE-native SKILL.md delivery) as one installable workflow spec — as of September 8, 2026.
  • Useful for a specific audience: Developers doing in-project research, competitive analysis, or architecture decisions where output needs to live alongside code and be auditable. Developer sentiment (HN/Reddit Aug-Sep 2026) confirms a hybrid preference: chat for exploration, files for durable truth. Low value for fast ad-hoc lookups. [stangeli.substack.com]
  • Genuinely unique combination: No alternative ships all six patterns — IDE artifact delivery, per-claim atomic confidence with full URLs, parallel DAG subagents, plan pre-mortem, self-scoring patch loop, and SKILL.md portability — as one installable spec. The Perplexity Agent API (replacing Sonar API Sep 27, 2026) is the closest gap-narrowing move, but still produces no atomic CLAIM/SOURCE/FRESHNESS/CONFIDENCE evidence file. [docs.perplexity.ai]
  • Individual patterns are academically validated (Flash-Searcher, PreFlect, DeepVerifier, LedgerMind, ArcticSwarm, DualStake) but no paper or tool combines them. A new Jul-Sep 2026 wave (Interaction Tax, SILO-BENCH) cautions that multi-agent gains are conditional on structured isolation — naive coordination can erase diversity. [arxiv.org]
  • Citation quality gap is real and persistent: Haus Research audit (Sep 2, 2026) found 34.7% of Perplexity citation markers fail a numerical accuracy check. DeepTRACE ranking: ChatGPT GPT-5 79.1% citation accuracy, Claude Opus 4.5 77%, Perplexity 58.0%, Gemini 40.3%. deep-research skill enforces a full URL per atomic claim — no alternative does this by default. [hausresearch.com]
  • Gemini Max added MCP, but with a regression: Gemini Deep Research Max launched MCP support and native chart generation (~160 searches/task), but a Sep 4, 2026 regression caused MCP tools to be listed but never invoked (fix promised Sep 7). Perplexity Agent API is the more durable structural threat — but both still lack the atomic evidence pipeline. [support.google.com]

Charts

Citation accuracy by tool (DeepTRACE + Haus, Sep 2026)

ChatGPT GPT-5 79.1%
Claude Opus 4.5 77%
Perplexity 58.0%
Gemini Deep Research 40.3%

Source: arxiv.org/DeepTRACE (Sep 2025 snapshot). Perplexity Haus audit (numeric claim accuracy): 34.7% failure rate — different methodology, conflicted with DeepTRACE.

Feature coverage: deep-research skill vs alternatives

Feature deep-research ChatGPT DR Perplexity Copilot /research GPT-Researcher
Runs inside IDE Yes No No CLI No
Atomic evidence store Yes No No No No
Per-claim confidence ratings 4-tier No No Section No
Parallel subagent DAG Explicit Unknown Unknown No LangGraph
Plan pre-mortem phase Yes No No No No
Self-scoring patch loop 0–12 No No No No
HTML dashboard output Yes No No No No
MCP tool integration Inherits host No API only Hardcoded No
SKILL.md portability 41 products No No CLI only No
Open source MIT No No No MIT

skills.sh deep-research install leaderboard (Sep 8, 2026)

firecrawl-deep-research 33,752
parallel-deep-research 13,807
199-biotechnologies 9,675
affaan-m/ecc 8,826
TheAdamLabs/deep-research-skill not listed

Source: skills.sh API · Sep 8, 2026

Key Findings

Consumer research landscape

Five major tools remain chat-bound with no structured per-claim evidence store. ChatGPT GPT-5 leads citation accuracy at 79.1% (DeepTRACE); Perplexity 58.0%; Gemini 40.3%. Haus audit (Sep 2, 2026): 34.7% of Perplexity numeric citations fail verification. Gemini Max adds ~160 searches/task and MCP, but Sep 4 regression undercuts it. [arxiv.org/DRBench]

SKILL.md ecosystem

agentskills.io lists 41 compliant products (Cursor, Claude Code, VS Code, Gemini CLI, GitHub Copilot). skills.sh: 9,728 skills, 1.39M all-time installs. TheAdamLabs/deep-research-skill has 1 GitHub star, no skills.sh listing, no tracked install count. [agentskills.io]

Competing IDE tools

Copilot CLI /research is the closest first-party competitor. Added HTML export (April 2026), but still hardcoded tools and no atomic CLAIM schema. web-researcher-mcp v1.49.3 (Sep 1) and openresearch-mcp provide structured research JSON — but leave orchestration to the calling agent. No Aug-Sep 2026 launch replicates the five-pattern stack. [docs.github.com]

Academic validation (Jul-Sep 2026)

ArcticSwarm and AREX confirm gated multi-agent isolation gains. Interaction Tax (Aug 2026): naive coordination erases diversity — gains conditional on structured architecture. DualStake (EMNLP 2026): calibrated per-claim confidence scoring. No paper combines all five patterns in one installable spec. PreFlect: +8.5pp on GAIA. DeepVerifier: +8-11%. [arxiv.org/Interaction Tax]

Sep 3-8 competitive delta: what changed in 5 days
  • Perplexity Sonar API retirement (Sep 27 deadline confirmed): sonar-deep-research maps to Agent API high/xhigh with typed search_results steps and MCP. Closest structural gap-narrowing move, but still no atomic evidence schema.
  • Gemini Deep Research Max MCP regression (Sep 4): tools listed but never invoked; Google promised backend fix by Sep 7. MCP + native charts in Max mode remain on paper.
  • Entire Agentic Search (Sep 3): CLI skill integration + self-scoring benchmark harness (Searchmark), but targets code-history search, not open-web research.
  • ChatGPT Deep Research: no product changes Sep 3-8.
  • Copilot CLI /research: unchanged on core limitations. Issue #4076 still open.
Developer workflow evidence: why file artifacts win for "durable truth"

HN and Reddit threads (Aug 2026) repeatedly describe chat history as linear, unindexed, and untraceable after a few weeks. Third-party tools (DeepCite, Universal Markdown Exporter) exist because native chat/Deep Research output loses citations. The dominant 2026 pattern is hybrid: chat for discovery and synthesis, file-based artifacts for verified output that compounds over time. FS-Researcher confirms file-system workspaces add negligible latency (<0.03% of wall-clock). [arxiv.org/FS-Researcher]

Evidence Confidence

Completeness
3/3
Accuracy
3/3
Relevance
3/3
Artifact fit
3/3
high 212 claims medium 52 claims low 2 claims conflicted 2 claims

268 total atomic claims across 8 subagents · 40+ searches

Confidence Claim Source Freshness
high 34.7% of Perplexity citation markers fail a numerical accuracy check hausresearch.com Sep 2, 2026
high ChatGPT GPT-5 Deep Research: 79.1% citation accuracy (DeepTRACE); Perplexity 58.0%; Gemini 40.3% arxiv.org/DeepTRACE Sep 2025 snapshot
high Claude Opus 4.5 achieves 77% factual claim accuracy (14-model study, May 2026) arxiv.org/2605.06635 May 2026
high Perplexity retiring Sonar API Sep 27, 2026; maps to Agent API high/xhigh with MCP docs.perplexity.ai Sep 2026
high Gemini Deep Research Max MCP regression Sep 4, 2026: tools listed but never invoked support.google.com Sep 4, 2026
high Gemini hallucinated URL rate: 13.3% (DRBench); OpenAI: 3.5%; Claude: 3.0-3.2% arxiv.org/2604.03173 Apr 2026
high Increased search depth degrades factual accuracy ~42% while link accessibility stays above 92% arxiv.org/2605.06635 May 2026
high PreFlect pre-mortem planning achieves +8.5pp on GAIA over Reflexion arxiv.org/2602.07187 Feb 2026
high DeepVerifier self-scoring loop delivers +8–11% accuracy on GAIA without training aclanthology.org ACL 2026 ACL 2026
high Flash-Searcher DAG parallel reduces steps 35% (11.2→7.4); 67.7% BrowseComp accuracy arxiv.org/2509.25301 Sep 2025 (ICLR 2026)
high Interaction Tax (Aug 2026): naive multi-agent interaction erases diversity; gains conditional on structured isolation arxiv.org/2508.14982 Aug 2026
high Copilot CLI /research: tool list is hardcoded, issue #4076 still open Sep 8, 2026 github.com/copilot-cli Sep 2026
high Copilot CLI /research added HTML export April 2026; still no atomic CLAIM schema docs.github.com 2026
high Microsoft retired consumer Copilot Deep Research on Aug 18, 2026 support.microsoft.com Aug 2026
high agentskills.io lists 41 SKILL.md compliant products as of Sep 2026 agentskills.io/clients Sep 2026
high firecrawl-deep-research leads skills.sh deep-research leaderboard at 33,752 installs skills.sh API Sep 8, 2026
high SKILL.md workflows ignored ~50% of the time when overlapping built-in model behaviors github.com/anthropics Mar 2026
high Multi-agent research uses ~15× more tokens than ordinary chat anthropic.com Jun 2025
high FS-Researcher: file-system workspace adds negligible latency (<0.03% of wall-clock) arxiv.org/2602.01566 2026
medium DualStake (EMNLP 2026): calibrated per-claim confidence scoring framework arxiv.org/2509.04821 Sep 2026
medium No Aug-Sep 2026 launch fully replicates deep-research skill's five-pattern stack synthesized Sep 2026
medium Developer sentiment (HN/Reddit Aug-Sep 2026): hybrid workflow — chat for exploration, files for durable truth pickuma.com 2026
medium TheAdamLabs/deep-research-skill: 1 GitHub star, no skills.sh listing github.com/theadamlabs Sep 2026
medium Gartner projects 40%+ of agentic AI initiatives cancelled by 2027 joeyhipolito.dev Mar 2026
conflicted Perplexity citation accuracy: 34.7% numeric failure (Haus) vs 58.0% accuracy (DeepTRACE) vs 90.24% (DeepResearch Bench) — different product surfaces and metrics hausresearch.com Sep 2026
conflicted Confidence tiers are prompt heuristics, not calibrated against ECE/AUC benchmark metrics arxiv.org/2601.15778 Jan 2026

Sources (top 25 by importance)

Gaps & Follow-up