Skip to content

LegionForge/Briarios

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

LegionForge-Briarios

Status: R&D

Named for Briareos (Βριάρεως) — one of the Hecatoncheires of Greek myth. 50 heads, 100 arms. An entity that holds and executes a hundred things simultaneously. In the Appleseed universe, the cyborg who protects and orchestrates.

Briarios is a security-native orchestration meta-layer for parallel agentic development workstreams.

It does not replace OpenHands, LangGraph, or Anthropic Managed Agents. It sits on top of them and adds what they don't have: security-first issue triage, model-tier routing, multi-resource budget gating, and an independent verification pipeline.


What Problem This Solves

Modern AI agent frameworks (CrewAI, AutoGen, LangGraph) give you the primitives to run parallel agents. What they don't give you:

  • Security-aware triage — not all issues have the same blast radius; a CVE and a UI label cleanup shouldn't compete for the same agent tier
  • Model-tier routing — routing Opus at a cosmetic refactor is waste; routing Haiku at a CVE is risk
  • Multi-resource budget gating — token budgets exist; CPU/GPU/VRAM awareness does not
  • Independent verification — no existing orchestrator enforces a separate pentest/eval agent reviewing every change before merge
  • Security-native guardrails — scope enforcement, no-direct-to-main, mandatory risk assessment per change

Architecture (Meta-Layer)

┌─────────────────────────────────────────────────────┐
│                  BRIARIOS META-LAYER                │
│                                                     │
│  ┌─────────────┐  ┌──────────────┐  ┌───────────┐ │
│  │  Triage     │  │  Model-Tier  │  │  Budget   │ │
│  │  Scorer     │→ │  Router      │→ │  Gate     │ │
│  └─────────────┘  └──────────────┘  └───────────┘ │
│                                           │         │
│                          ┌────────────────┘         │
│                          ▼                          │
│  ┌─────────────────────────────────────────────┐   │
│  │           AGENT EXECUTION LAYER             │   │
│  │  OpenHands (coding) │ LangGraph (workflows) │   │
│  │  Anthropic Managed Agents (orchestration)   │   │
│  └─────────────────────────────────────────────┘   │
│                          │                          │
│                          ▼                          │
│  ┌─────────────────────────────────────────────┐   │
│  │         VERIFICATION PIPELINE               │   │
│  │  Test Agent → Review Agent → Eval Agent     │   │
│  │  → Pentest Agent (security tier only)       │   │
│  └─────────────────────────────────────────────┘   │
│                          │                          │
│                   Human Gate (required)             │
└─────────────────────────────────────────────────────┘

Triage Scoring

Every issue is scored on four axes before assignment:

Axis Low (1) Med (3) High (5)
Security risk cosmetic functional gap exploitable CVE
Blast radius isolated file subsystem cross-cutting
Complexity unambiguous some judgment ADR needed
Blocking standalone blocks 1 blocks 2+

Tier assignment:

  • S (score 16–20): Security issues → Opus
  • A (score 10–15): Complex bugs, ADRs → Sonnet
  • B (score 5–9): Simple bugs → Sonnet/Haiku
  • C (score 1–4): Quick wins, cosmetic, docs → Haiku / local model

Budget Gate

Before spawning any parallel agent lane:

  • Active agent count < MAX_PARALLEL (default: 3)
  • Token spend today < DAILY_TOKEN_BUDGET
  • CPU load < 80%
  • GPU VRAM headroom > MIN_VRAM_GB
  • GitHub API requests < rate limit threshold

All conditions must pass. Any failure queues the agent and retries on next budget tick.


Verification Pipeline

Agent produces diff
  → Test Agent        (runs CI, reports pass/fail)
  → Review Agent      (reads diff, checks scope creep + security)
  → Eval Agent        (checks diff against issue Done-When criteria)
  → Pentest Agent*    (*security tier only — checks fix doesn't introduce new vulns)
  → Human gate        (no merge without explicit approval)

Each agent in the pipeline is independent — it sees only the artifact, not the prior agent's reasoning.


What We Build vs. What We Consume

Capability Source
Coding agent execution OpenHands (open source, 65K+ stars)
Workflow graph orchestration LangGraph (already in LegionForge)
Managed multi-agent API Anthropic Managed Agents (public beta)
Triage scorer Briarios — build
Model-tier router Briarios — build
Multi-resource budget gate Briarios — build
Verification pipeline Briarios — build
Security-native guardrails Briarios — build

License

MIT — see LICENSE

Architectural Decisions

All four open questions from initial research are now resolved:

ADR Decision
ADR-001 Bolt-on meta-layer — consume OpenHands + LangGraph + Anthropic Managed Agents
ADR-002 LangGraph-native for S/A tier; OpenHands for B/C tier
ADR-003 Polling (v0.1) → event-driven Prometheus/Alertmanager (v0.2+)
ADR-004 Sequential with early exit — fail fast, 1–2x token cost vs 3–4x for parallel
ADR-005 Hybrid: GitHub Projects (display) + Postgres (orchestration state)

Status

Research complete. ADRs accepted. Ready for scaffolding.

See docs/research/build-vs-buy.md for full landscape analysis.

About

Security-native meta-layer for parallel agentic development workstreams — triage scoring, model-tier routing, budget gating, and verification pipelines on top of OpenHands + LangGraph

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors