raven-multi-agent-orchestrator

name

raven-multi-agent-orchestrator

description

Orchestrate 100+ specialized AI agents across an ensemble of frontier and distilled models to discover, debate, validate, prove, enrich, and triage exploitable bugs end-to-end. Use when the user wants to run the full MDASH pipeline with multi-model ensemble, dynamic agent spawning, or production-scale vulnerability discovery. Every finding terminates at a tool oracle (𝒯), classical-ML detector (𝓜), or scored hypothesis (𝓛), and ships with D3FEND defensive techniques, ATT&CK offensive context, exposure scoring, and DevSecOps routing.

Raven — Multi-Agent Orchestrator

This skill is the master conductor of the MDASH (Multi-Model Agentic Security Harness) pipeline. It spawns, routes, and manages 100+ specialized agents across 7 pipeline stages, dynamically allocating frontier models (GPT-4o, Claude 3.5) for complex reasoning and distilled models for high-volume tasks.

Built for production DevSecOps at enterprise scale:

• Every finding has an owner, a triage process, and a Patch Tuesday deadline
• 100+ distinct agent classes for maximum type safety and explicit behavior
• Dynamic model allocation by task complexity — no waste, no compromise
• Parallel async execution with built-in rate limiting
• Collaboration between ACS (Autonomous Code Security), MORSE (Offensive Research & Security Engineering), and WARP (Windows Attack Research and Protection)

When to use

Trigger when the user asks any of:

• "Run the full MDASH pipeline"
• "Orchestrate multi-agent security scanning"
• "Discover vulnerabilities at scale"
• "Run 100+ agents on this codebase"
• "Multi-model debate and validation"
• "Production vulnerability discovery"
• "Enterprise-scale security audit"

Do not trigger for: single-agent tasks (use individual auditor agents), defensive actions only (use raven-zero-day-defend), or incident response (use raven-zero-day-investigator).

Inputs

Required:

1. Target path — path to codebase or repository
2. Target type — c-cpp-source, solana-program, evm-contract, rust-source, windows-kernel, hyper-v, azure-service
3. Budget — max USD spend (default $10.00)

Optional:

• Stages — subset of stages to run (default: all 7)
• Debate panel size — 3-7 debaters (default 3)
• Vote threshold — 0.0-1.0 for validation (default 0.67)
• Max concurrent — parallel agent limit (default 50)

Pipeline — seven mandatory stages

Stage 1 — PREPARE (𝒯-grounded)

Spawn prepare agents (CodebaseIngesterAgent, ThreatModelerAgent, LanguageIndexerAgent) to:

1. Ingest and parse the target codebase
2. Build a ThreatModel identifying high-risk areas and trust boundaries
3. Index languages and build cross-reference maps
4. Output: SurfaceMap with annotated risk areas

Stage 2 — SCAN (𝒯 + 𝓜 + 𝓛)

Spawn 40+ auditor agents by bug class. Each agent is a distinct class with domain-specific instructions:

• MemoryCorruptionAuditor — buffer overflows, UAF, double-free
• HeapExploitationAuditor — heap spray, chunk manipulation
• StackCorruptionAuditor — stack canary bypass, ROP gadgets
• RaceConditionAuditor — TOCTOU, double-fetch, synchronization errors
• AuthBypassAuditor — missing checks, privilege escalation
• WindowsKernelAuditor — IRP flaws, lock invariants (private codebase reasoning)
• HyperVAuditor — VM escape, hypercall flaws
• EVMReentrancyAuditor — Solidity reentrancy, CEI pattern violations
• SolanaAccountConfusionAuditor — missing signer, PDA validation
• ... (40+ more distinct auditors)

Dynamic model allocation: frontier models for complex kernel/driver analysis, distilled models for simple pattern matching.

Output: CandidateFinding records with bug_class, location, precondition, severity, and falsification test.

Stage 3 — VALIDATE (𝓛-scored, multi-model debate)

For each candidate finding, spawn a debate panel of 3-7 debater agents with different models and personas:

• ProVulnerabilityDebater — argues FOR exploitability
• AntiVulnerabilityDebater — argues AGAINST (challenges evidence)
• ArbiterDebater — neutral evaluation and confidence scoring
• TechnicalDebater — focuses on technical feasibility
• ImpactDebater — evaluates real-world impact

Each debater uses a different model (Claude 3.5, GPT-4o, Gemini) to maximize independent reasoning. Votes are aggregated with configurable threshold (default 67% FOR to validate).

Output: ValidatedFinding with debate transcript, confidence score, and validation status.

Stage 4 — DEDUP (𝓜)

Spawn dedup agents (EmbeddingGeneratorAgent, ClusteringEngineAgent) to:

1. Generate embeddings for validated findings
2. Cluster by semantic similarity (DBSCAN)
3. Merge duplicate findings, preserving highest confidence variant

Output: DeduplicatedFinding records with cluster size and merge history.

Stage 5 — PROVE (𝒯-grounded)

Spawn prover agents to generate and validate proof-of-concepts:

• PoCGeneratorAgent — constructs triggering inputs
• HarnessBuilderAgent — builds libFuzzer/AFL++/honggfuzz harnesses
• SanitizerRunnerAgent — runs ASan/TSan/UBSan
• FormalVerificationProverAgent — SMT solver + symbolic execution

Output: ProvenFinding with sanitizer reports, harness paths, and proof status.

Stage 6 — ENRICH (𝒯-grounded)

Spawn enricher agents to bind full threat context:

• D3FENDEnricherAgent — maps CWE to real 271-entry MITRE D3FEND catalog
• ATTACKMapperAgent — binds offensive ATT&CK techniques
• ExposureScorerAgent — calculates real-world exposure 0-100
• ComplianceMapperAgent — CCI-to-NIST SP 800-53 mappings

Output: Enriched findings with d3fend_techniques, attack_techniques, exposure_score, compliance_controls, and threat_narrative.

Stage 7 — TRIAGE (𝒯-grounded)

Spawn triage agents for DevSecOps routing:

• SeverityTriage — CVSS-based severity scoring
• OwnerAssignment — assigns finding owner by component responsibility
• PatchTuesdayScheduler — schedules fix deadlines
• FalsePositiveFilter — filters probable FPs before human review

Output: Final report with finding_owner, team, patch_tuesday_target, sla_hours, and recommended_action.

CDP contract — quick check before returning

assert every finding terminates at 𝒯 or 𝓜 or 𝓛
assert every finding has d3fend_techniques from pipeline.d3fend.enrichment_engine
assert every finding has attack_techniques from pipeline.d3fend.attack_mapper
assert every finding has exposure_score from pipeline.threat_intel.onchain_threat_intel
assert every finding has finding_owner for DevSecOps routing
assert every finding has patch_tuesday_target
assert every finding has sla_hours
assert report.confidence == "high" only if debate confidence >= 0.67 AND reproducible=true
assert no stage silently skipped without gap_warning

Output contract

# MDASH Multi-Agent Report — <target>

## Summary
- Target: <path> (<type>)
- Wall time: <s>
- Total cost: $<usd>
- Findings: <n>
- Agents spawned: <n>
- Models used: <list>

## Findings

### Finding F-<n>
- ID: <id>
- Bug class: <class>
- Location: <file:line>
- Severity: <critical|high|medium|low>
- Confidence: <float>
- Validated: <yes|no>
- Proven: <yes|no>
- D3FEND: <technique list>
- ATT&CK: <technique list>
- Exposure: <score>/100
- Owner: <email>
- Patch Tuesday: <date>
- SLA: <hours>
- Recommended action: <route-to-investigator|route-to-defend>
- Threat narrative: <paragraph>
- Debate transcript: <link>
- Proof of concept: <path>

(repeat)

## Stage metrics
| Stage | Agents | Findings | Wall time | Cost |
|-------|--------|----------|-----------|------|
| prepare | <n> | <n> | <ms> | $<usd> |
| scan | <n> | <n> | <ms> | $<usd> |
| validate | <n> | <n> | <ms> | $<usd> |
| dedup | <n> | <n> | <ms> | $<usd> |
| prove | <n> | <n> | <ms> | $<usd> |
| enrich | <n> | <n> | <ms> | $<usd> |
| triage | <n> | <n> | <ms> | $<usd> |

## Agent metrics
| Agent | Tasks | Cost | State |
|-------|-------|------|-------|
| ... |

## Budget status
- Budget: $<usd>
- Spent: $<usd>
- Remaining: $<usd>
- Utilization: <percent>%

## Reproducibility kit
- Commit: <sha>
- Config: <path>
- Agent definitions: <count>
- Model mappings: <path>

Refusal rules

1. Refuse to run without budget cap configured.
2. Refuse to skip validation stage — every finding must pass debate.
3. Refuse to emit findings without D3FEND + ATT&CK enrichment.
4. Refuse to bypass triage — every finding must have an owner.
5. Refuse to run on unauthorized targets.
6. Refuse to exceed max concurrent agents without explicit override.

Related skills and modules

• raven-zero-day-hunter — upstream skill for initiating hunts
• raven-zero-day-detection — downstream alert producer
• raven-zero-day-investigator — downstream for alert investigation
• raven-zero-day-defend — downstream for defensive action
• pipeline/multi_agent_orchestrator/ — source-of-truth implementation
• pipeline/openai_agents_adapter/ — optional OpenAI Agents SDK backend
• pipeline/eval/ — benchmark evaluation framework
• pipeline/d3fend/ — D3FEND + ATT&CK enrichment engine
• pipeline/threat_intel/ — on-chain exposure scoring