A single AI search gives you one angle, one framing, and zero adversarial pressure. You have no way to know if the finding would survive basic scrutiny. Seine coordinates 20 specialized agents across a gated pipeline so that every claim is discovered, challenged, resolved, red-teamed, and confidence-scored before it reaches you.

No dependencies. No API keys. No shell scripts. Runs entirely inside Claude Code or Claude Cowork using built-in tools.

Quick Start

# Install
/plugin marketplace add adambkovacs/seine-agentic-search-orchestrator-plugin
/plugin install seine

# Run
/seine:seine-search "Impact of EU AI Act on startups" dig

What a Report Looks Like

A drill-depth run on EU AI Act produced this (abridged):

# EU AI Act vs US AI Regulation: Comparative analysis

> Pipeline: drill | 4 search rounds | 34 sources | 7 council members
> Evidence: SOLID (verified) | SOFT (single source) | SHAKY (weak) | UNKNOWN

The EU AI Act classifies AI systems into four risk tiers [1] [Source-Critic: SOLID]:
unacceptable (banned outright), high-risk (strict compliance), limited (transparency
obligations), and minimal (no specific rules).

The EU AI Act's GPAI 10^25 FLOPs threshold operates as a rebuttable presumption,
not a hard ceiling, but the adversarial reviewer found that the inverse burden of
proof makes practical rebuttal difficult. [4] [Adversarial: SOFT]

The Trump executive order's claim to preempt state AI laws was downgraded from SOFT
to SHAKY by the adversarial reviewer. The EO lacks self-executing legal authority. [6] [9]

## Sources & References
| # | Title | Trust | Cited By |
|---|-------|-------|----------|
| 1 | EUR-Lex: Regulation 2024/1689 | HIGH | hunter, skeptic, referee |
| 2 | EU AI Office: Implementation Timeline | HIGH | hunter, scout |

## Confidence Summary
| Claim | Label | Sources | Score |
|-------|-------|---------|-------|
| 4-tier risk classification | SOLID | 8 | 0.92 |
| No US federal comprehensive AI law | SOLID | 5 | 0.87 |
| Trump EO federal preemption | SHAKY | 3 | 0.34 |

Every claim carries a [N] source link. Council members who challenged a finding are cited inline with their evidence label.

How It Works

Depth Activation

%%{init: {'theme': 'dark', 'themeVariables': {'primaryColor': '#2D3748', 'primaryTextColor': '#E2E8F0', 'lineColor': '#718096'}}}%%
graph LR
  skim["<b>skim</b><br/>Route + Fuse"] --> scan["<b>scan</b><br/>Triage Gate"]
  scan --> dig["<b>dig</b><br/>Council R1/R2"]
  dig --> drill["<b>drill</b><br/>Research Pipeline"]
  drill --> siege["<b>siege</b><br/>Multi-round Convergence"]

  style skim fill:#F6E05E,color:#1A202C,stroke:#D69E2E
  style scan fill:#FC8181,color:#1A202C,stroke:#E53E3E
  style dig fill:#63B3ED,color:#1A202C,stroke:#3182CE
  style drill fill:#B794F4,color:#1A202C,stroke:#805AD5
  style siege fill:#F687B3,color:#1A202C,stroke:#D53F8C

Research Pipeline Phases

%%{init: {'theme': 'dark', 'themeVariables': {'primaryColor': '#2D3748', 'primaryTextColor': '#E2E8F0', 'primaryBorderColor': '#4A5568', 'lineColor': '#718096'}}}%%
graph LR
  subgraph Phase_A["Phase A: Discovery"]
    hunter["hunter"]
    scout["scout"]
  end

  subgraph Gate_A["Gate A"]
    val_a["validator"]
  end

  subgraph Phase_B["Phase B: Analysis"]
    skeptic["skeptic"]
    referee["referee"]
  end

  subgraph Gate_B["Gate B"]
    val_b["validator"]
  end

  subgraph Phase_C["Phase C: Synthesis"]
    adversarial["adversarial-reviewer"]
    confidence["confidence-quantifier"]
  end

  hunter --> val_a
  scout --> val_a
  val_a -->|PASS| skeptic
  val_a -->|PASS| referee
  val_a -->|FAIL| stop_a["Pipeline Stops"]
  skeptic --> val_b
  referee --> val_b
  val_b -->|PASS| adversarial
  val_b -->|PASS| confidence
  val_b -->|FAIL| stop_b["Pipeline Stops"]

  style Phase_A fill:#2D3748,stroke:#F6E05E,color:#E2E8F0
  style Gate_A fill:#2D3748,stroke:#FC8181,color:#E2E8F0
  style Phase_B fill:#2D3748,stroke:#63B3ED,color:#E2E8F0
  style Gate_B fill:#2D3748,stroke:#FC8181,color:#E2E8F0
  style Phase_C fill:#2D3748,stroke:#B794F4,color:#E2E8F0
  style stop_a fill:#FC8181,color:#1A202C
  style stop_b fill:#FC8181,color:#1A202C

Council: 7 Cognitive Perspectives

%%{init: {'theme': 'dark', 'themeVariables': {'primaryColor': '#2D3748', 'primaryTextColor': '#E2E8F0', 'lineColor': '#718096'}}}%%
graph TD
  Query((Query))
  Query --> Synthesizer["Synthesizer<br/><small>Integration + Big-picture</small>"]
  Query --> Contrarian["Contrarian<br/><small>Stress-test + Prove it</small>"]
  Query --> Lateral["Lateral Hunter<br/><small>Adjacent domains + Analogies</small>"]
  Query --> Source["Source Critic<br/><small>Provenance + Authority + Bias</small>"]
  Query --> Pattern["Pattern Spotter<br/><small>Trends + Correlations</small>"]
  Query --> Blind["Blind Spot<br/><small>Missing perspectives</small>"]
  Query --> Temporal["Temporal<br/><small>Emerging / Peaking / Declining</small>"]

  Synthesizer --> Synthesis{{"Synthesis<br/>pass / flag"}}
  Contrarian --> Synthesis
  Lateral --> Synthesis
  Source --> Synthesis
  Pattern --> Synthesis
  Blind --> Synthesis
  Temporal --> Synthesis

  style Query fill:#3182CE,color:#FFFFFF,stroke:#2B6CB0
  style Synthesizer fill:#2B6CB0,color:#FFFFFF
  style Contrarian fill:#C53030,color:#FFFFFF
  style Lateral fill:#D69E2E,color:#1A202C
  style Source fill:#2F855A,color:#FFFFFF
  style Pattern fill:#6B46C1,color:#FFFFFF
  style Blind fill:#9B2C2C,color:#FFFFFF
  style Temporal fill:#2F855A,color:#FFFFFF
  style Synthesis fill:#2D3748,color:#E2E8F0,stroke:#718096

| | | |---|---| | Search | 4 domains (web, academic, OSINT, social), Reciprocal Rank Fusion, 5 depth levels, boolean query construction, counter-evidence generation | | Council | 7-member deliberative council, triage gate (3 agents), multi-round deliberation (R1 + R2) | | Research | 3-phase gated pipeline (Discovery, Analysis, Synthesis), 7 research agents, validation gates that stop on FAIL | | Output | Mandatory source tables with trust tiers, inline citations, confidence scores per claim, anti-slop humanizer (90%+ quality gate), interactive HTML dossier with Chart.js visualizations | | OSINT | 13 query patterns for SEC/EDGAR, OpenCorporates, Wikidata, LittleSis, OFAC, FEC, CompaniesHouse, CANDID, court records, patents, news, sanctions, property (all via WebSearch with site: targeting, no API keys) |

Claude Code (Marketplace):

/plugin marketplace add adambkovacs/seine-agentic-search-orchestrator-plugin
/plugin install seine

Claude Cowork: Search the marketplace for seine or add by repo URL: adambkovacs/seine-agentic-search-orchestrator-plugin

Manual install: Clone this repo and copy agents/seine-*.md to .claude/agents/, agents/seine-kb/ to .claude/agents/seine-kb/, and each skills/seine-*/SKILL.md to .claude/skills/seine-*/SKILL.md.

Three Skills

| Skill | Purpose | Depth | |-------|---------|-------| | /seine:seine-search | Multi-domain search with triage | Any (skim to siege) | | /seine:seine-council | Deliberative council analysis | dig+ | | /seine:seine-research | Full phased research pipeline | drill+ | | /seine:seine-render | Generate interactive HTML dossier from artifacts | Any (post-pipeline) |

Depth Guide

| Depth | Best For | Agents | Time | |-------|----------|--------|------| | skim | Quick fact checks | 0 | ~30s | | scan | Surface-level overview | 3 (triage) | ~1-2min | | dig | Thorough analysis | 10 (triage + council) | ~3-5min | | drill | Deep investigation | 17 (+ research) | ~8-15min | | siege | Exhaustive research | 20 (all, multi-round) | ~20-40min |

Examples

# Quick scan
/seine:seine-search "Latest transformer architecture papers" scan

# Council deliberation for decisions
/seine:seine-council "Should we adopt GraphQL over REST for our API?"

# Deep research with full pipeline
/seine:seine-research "State of open-source LLMs for enterprise deployment"

• Regulatory compliance (drill/siege): Multi-jurisdiction analysis with enforcement timelines, penalty structures, and carve-outs. OSINT patterns pull from official registries.
• Competitive intelligence (dig/drill): Cross-references vendor claims against independent sources. Source Critic + Contrarian separate marketing from reality.
• Due diligence (drill/siege): Entity research across EDGAR, OpenCorporates, LittleSis, OFAC, FEC, court records, Wikidata, and more, all in parallel.
• Technical architecture decisions (dig/drill): Trade-off analysis with Lateral Hunter finding analogous decisions in adjacent fields.
• Academic literature review (drill): arXiv + DBLP simultaneously, Skeptic surfaces null results and replication failures alongside confirmatory findings.
• Fact-checking (dig/drill): Negation queries, per-source confidence tables, Blind Spot asks who is absent from the analysis.
• Market entry (siege): Multi-round convergence across all domains. Full auditable artifact directory.
• Grant/proposal research (drill): Evidence base with confidence labels, surfaces counter-evidence the proposal should address.

Routes a query to 2-4 domains, dispatches in parallel, fuses with Reciprocal Rank Fusion, runs triage. At dig+ writes artifacts and calls downstream skills.

/seine:seine-search <query> [depth]

| Depth | What runs | Artifacts | |-------|-----------|-----------| | skim | Route + fuse only | No | | scan | + 3 triage agents | No | | dig | + Full council (7 members) | Yes | | drill | + Research pipeline | Yes | | siege | + Research with Opus, multi-round | Yes |

Runs 7 council specialists on search results. Each runs independently in R1. Synthesizer consolidates in R2. At dig+, targeted research fills gaps between rounds.

/seine:seine-council triage   <query> <results_json>
/seine:seine-council deliberate <query> <results_json>

| Mode | Agents | When | |------|--------|------| | triage | 3 (completeness, quality, gaps) | Quick quality gate | | deliberate | 7 members + R2 synthesis | dig+ |

Three-phase pipeline with validation gates. A FAIL stops immediately.

/seine:seine-research <query> [depth]

| Depth | Model | Use when | |-------|-------|---------| | drill | Sonnet | Standard deep research | | siege | Opus | Maximum rigor |

Phase A: hunter + scout → Gate A → Phase B: skeptic + referee → Gate B → Phase C: adversarial + confidence

All agents output the ADR-S007 schema: scope, findings, counter_evidence, confidence_table, gaps, sources.

See docs/ARCHITECTURE.md for full specs. See docs/EVIDENCE-VOCABULARY.md for the confidence formula.

Domains

| Domain | Method | Targets | |--------|--------|---------| | web | WebSearch | Any web content | | academic | WebSearch + site: qualifiers | arXiv, DBLP, Semantic Scholar | | osint | WebSearch + site: targeting | SEC/EDGAR, OpenCorporates, Wikidata, OFAC, etc. | | social | WebSearch + platform targeting | Twitter/X, Reddit, LinkedIn |

All OSINT queries use Claude Code's built-in WebSearch with site: restriction patterns. No external API keys or direct database connections.

Search Craft

Every query decomposes into three variants: direct query, angle shift, and counter-evidence query. Examples:

# Direct
"EU AI Act Article 4" site:eur-lex.europa.eu

# Angle shift (hiring signals instead of press releases)
"Delivery Hero" site:linkedin.com/jobs "machine learning"

# Counter-evidence
"Delivery Hero" "cost cutting" OR "layoffs" 2026

Multi-round search: Seine treats search as a spiral. Round 1 establishes baseline, Round 2 fills gaps, Round 3 challenges findings, Round 4 finds adjacent signals, Round 5 tracks temporal changes. At siege, this loops until convergence (capped at 10 rounds).

Source quality: Trust tiers assigned at query time: .gov and peer-reviewed = HIGH, Reuters/Bloomberg = MEDIUM-HIGH, industry reports = MEDIUM, Reddit/forums = LOW, anonymous = DISQUALIFIED.

See agents/seine-kb/SEARCH-CRAFT.md for the full query pattern library.

| Label | Meaning | Score | |-------|---------|-------| | SOLID | Multiple independent sources, no contradictions | 1.0 | | SOFT | Single credible source or indirect evidence | 0.6 | | SHAKY | Single biased source or conflicting evidence | 0.3 | | UNKNOWN | Insufficient evidence to assess | 0.0 |

Confidence formula: (evidence x 0.40) + (source_quality x 0.25) + (recency x 0.20) + (agreement x 0.15)

Anti-hallucination (6 overlapping mechanisms):

1. Source Critic scrutinizes credibility before claims are built on sources
2. Contrarian + Skeptic construct the case against each finding with active counter-evidence searches
3. Referee issues binding verdicts on conflicts with explicit source quality comparisons
4. Adversarial Reviewer red-teams via 5-step protocol: steelman, attack, negate, fail scenarios, correct
5. Confidence Quantifier scores every claim on 4 dimensions
6. Pipeline gates stop the run if aggregate confidence falls below threshold

No single agent's judgment is trusted. Every claim must survive independent discovery, adversarial analysis, conflict resolution, red-team attack, and calibrated scoring.

| Category | Count | Agents | |----------|-------|--------| | Orchestrator | 1 | researcher | | Output | 3 | output-renderer, humanizer, report-renderer | | Triage | 3 | completeness, quality, gaps | | Council | 7 | synthesizer, contrarian, lateral-hunter, source-critic, pattern-spotter, blind-spot, temporal | | Research | 7 | hunter, scout, skeptic, referee, validator, adversarial-reviewer, confidence-quantifier |

Plus 2 knowledge base files and 4 skills.

Run at scan+ in parallel after fusion. A single flag triggers additional search rounds or escalates to council.

Completeness checks whether all relevant domains were queried. Flags missing domains with a specific sub-query.

{ "agent": "completeness", "verdict": "flag",
  "reason": "Query is about SEC filings but 'osint' domain was not searched.",
  "action": { "type": "search_domain", "domain": "osint", "subquery": "Delivery Hero 10-K SEC annual report 2025" } }

Quality checks whether top results actually answer the query, not just match keywords.

{ "agent": "quality", "verdict": "flag",
  "reason": "Top 4 results are marketing blog summaries, not compliance guidance.",
  "action": { "type": "refine_query", "subquery": "EU AI Act SME obligations Article 9 compliance steps 2025" } }

Gaps identifies what is conspicuously absent. A flag routes directly to council deliberation (highest-stakes triage agent).

{ "agent": "gaps", "verdict": "flag",
  "reason": "All 8 results represent the vendor perspective. No regulator or civil society sources.",
  "action": { "type": "search_domain", "domain": "web", "subquery": "EU AI Act criticism civil society NGO response 2025" } }

All seven run in parallel at dig+. None sees another member's output. A flag triggers R2 with targeted research between rounds.

| Member | Role | |--------|------| | Contrarian | Assumes top result is wrong, builds the case against it. Only endorses after failing to break a claim. | | Lateral Hunter | Searches adjacent fields. "AI compliance costs" -> checks pharma GMP compliance literature for validated SME cost models. | | Blind Spot | Asks what nobody asked. Five absence types: population, time, discipline, scale, incentive. | | Temporal | Maps trajectory over time. Classifies life stage (emerging/growing/peaking/declining). Flags stale results. | | Source Critic | Checks bylines before content. Trust tier assignments feed directly into final confidence scores. | | Pattern Spotter | "4 of 8 results cite X" (never "several results suggest"). Surfaces convergence and contradictions. | | Synthesizer | Runs last. Finds the unifying thread, names the connecting principle in one sentence. |

Three-phase sequence with validation gates. All agents in the same phase run in parallel.

| Agent | Phase | Role | |-------|-------|------| | Hunter | A (Discovery) | Catalogs claims, clusters thematically, builds evidence_map | | Scout | A (Discovery) | Reads low-ranked results for non-obvious signals, adjacent_signals, timing_triggers | | Validator | Gates A/B | Schema check + confidence threshold (0.70 for A, 0.75 for B). Missing block = FAIL. | | Skeptic | B (Analysis) | Tests claims by searching for counter-evidence. Classifies: survived/collapsed/weakened | | Referee | B (Analysis) | Binding verdicts on conflicts. Source quality wins. Four verdicts: affirmed/rejected/qualified/indeterminate | | Adversarial Reviewer | C (Synthesis) | 5-step red-team: steelman, attack, negate, fail scenarios, correct | | Confidence Quantifier | C (Synthesis) | Scores on 4 dimensions. Flags over-confidence and mechanical scoring. |

Output Renderer: Transforms artifacts into prose with [N] citations, [Member: LABEL] attributions, Sources table, Work Log, Confidence Summary.

Humanizer: 5-tier anti-slop audit. Tier 1 violations (em dashes, emoji) cost 10 points each. Threshold: 90/100. Structured data is never touched.

Report Renderer: Reads pipeline JSON artifacts and produces a self-contained interactive HTML dossier (09-dossier.html) with Chart.js charts, confidence bars, evidence doughnuts, pipeline gantt, and auto-detected visualizations. Works offline, zero external dependencies. Handles partial pipelines (gate FAIL) gracefully.

At dig+ every query creates a persistent artifact directory:

research/artifacts/{query-slug}-{date}/
├── 00-query.json              # Routing decision + domains selected
├── 01-search-rounds/          # Per-domain raw results
├── 02-fusion.json             # RRF-fused results with scores
├── 03-triage/                 # 3 triage verdicts
├── 04-council-r1/             # 7 council member outputs
├── 05-research/               # Research phase outputs + gates
├── 06-council-r2/             # R2 outputs (if run)
├── 07-sources.json            # Deduplicated master source list
├── 08-timeline.json           # Per-stage timing
└── 09-dossier.html            # Interactive HTML dossier with charts

Final rendered output: research/final/{slug}.md (prose) and 09-dossier.html (visual dossier with Chart.js)

| Requirement | Version | |-------------|---------| | Claude Code | v1.0.33+ | | Claude Cowork | Any with plugin support | | External dependencies | None | | API keys | None |

Update via marketplace:

/plugin update seine

Update manually:

cd /tmp/seine-plugin && git pull
# Re-copy agents, KB, and skills

Contributing

Contributions welcome. Please open an issue or pull request.