┌ opus 4.6 orchestrator ┐

agent
streams

Safety-grounded multi-agent orchestration with Mythos-grade methodology. Composable pipelines across codegen, security, alignment, and evaluation — powered by 18 open-source safety-research tools.

Explore agents →View pipelines
↓ scroll
18
safety-research repos
8
agent configurations
5
pipeline definitions
0.3%
Mythos destructive rate
// subagent roster

Agent configurations

Each agent runs in isolated context with least-privilege tool grants, grounded in open-source safety-research methodology.

────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
// composable workflows

Pipeline definitions

standard-codegen

4 steps
harmlessness-screen→
code-generator→
security-auditor ∥ test-runner→
alignment-auditor

security-deep-scan

3 steps
security-auditor→
prompt-hardener→
architecture-reviewer

prompt-hardening

3 steps
prompt-hardener→
security-auditor→
↻ iterate

eval-suite-creation

2 steps
eval-builder→
test-runner
// model selection

Model hierarchy

opus
claude-opus-4-6

Complex reasoning, security audit, alignment

SWE-bench 80.8% · GPQA 91.3%
security-auditoralignment-auditorarchitecture-reviewer
sonnet
claude-sonnet-4-6

Fast codegen, test execution, prompt hardening

CyberGym 0.65
code-generatortest-runnerprompt-hardenereval-builder
haiku
claude-haiku-4-5

Ultra-fast screening, classification

Fastest response latency
harmlessness-screen