Stable names, shifting models. Declare intent — the matrix resolves the rest.
Agents hardcoded provider/model names. When upstream models changed or retired, entire agent fleets silently broke.
Copilot's claude-opus-4.6-fast costs 30× premium requests. Glob patterns like claude-opus-* silently matched it.
No way to express "for coding tasks, Sonnet outperforms GPT-5-mini" — providers were isolated silos.
Model rankings shift as new benchmarks drop. Yesterday's best is today's second-tier. Manual tracking across 66 agents doesn't scale.
Role names last years. Model assignments change weekly.
When a new model ships, update one matrix row — every agent benefits immediately.
Agents declare what KIND of work they're doing. The matrix resolves the best available model.
general catch-all
fast quick utility tasks
coding implementation
ui-coding frontend / spatial
security-audit vulnerability analysis
reasoning deep analysis
critique finding flaws
creative design direction
writing long-form content
research investigation
vision understanding visual input
image-gen generating images
critical-ops high-reliability infrastructure tasks
| Matrix | Philosophy | Best For |
|---|---|---|
| balanced (default) | Quality/cost sweet spot | Daily development |
| economy | Budget-first, free tiers preferred | Experimentation, bulk runs |
| quality | Best model everywhere | Critical production work |
| anthropic | Claude models only | Claude-centric shops |
| openai | OpenAI models only | OpenAI-centric shops |
| gemini | Gemini models only | Google ecosystem |
| copilot | Copilot-optimized, avoids 30× trap | GitHub Copilot users |
Curated using two leading independent evaluation sources.
Intelligence Index v4.0 — Terminal-Bench, SciCode, GPQA Diamond, 10+ evaluations
14 operational task categories, updated weekly
| Model | Highlight | Source |
|---|---|---|
| Gemini 3.1 Pro Preview | #1 overall — 57.18 Intelligence Index | Artificial Analysis |
| GPT-5.3 Codex | Top OpenAI coding model — "workhorse" | Weather Report |
| Claude Sonnet 4.6 | "May overtake Opus for everyday use cases" | Weather Report |
| Gemini 3 Flash | 46.43 index at $0.50/M — 218 tok/s | Artificial Analysis |
model_role: [coding, general]Hook module + YAML data files. No Amplifier Core changes required.
Works without the bundle installed — agents fall back to direct model config.
Available roles injected into context so LLMs can intelligently delegate to the right role.
List multiple roles in priority order. If security-audit can't resolve, try critique, then general.
When an LLM delegates to a sub-agent, it can override the role based on the specific task — e.g., sending image work to vision.
Existing agent configs work unchanged. Adding model_role is purely additive.
base keyword"Keep the matrix defaults, but try my preferred model first." Overrides compose with — not replace — the curated rankings.
Point coding to a local Ollama/vLLM endpoint. If it's down, the matrix seamlessly falls through to cloud providers.
Build and store custom matrices at ~/.amplifier/routing/ for team or project-specific strategies.
Before: claude-opus-* globs could match the 30× fast variant. Now: the matrix explicitly excludes it.
| Model | Copilot Multiplier | Routing Matrix Assignment |
|---|---|---|
| GPT-4.1 | 0× (free) | Not used — Haiku is better |
| Claude Haiku 4.5 | 0.33× | Fast roles |
| Claude Sonnet 4.6 | 1× | Coding / general |
| Claude Opus 4.6 | 3× | Reasoning / research only |
| Claude Opus 4.6 fast | 30× | NEVER — excluded from all matrices |
One glob-match accident at 30× premium can burn through a month of Copilot quota in a single session.
Combined setup dashboard — providers + routing in one flow
Interactive provider management with status display
Interactive matrix selection with live preview of role→model mappings
See exactly what each role resolves to right now
Switch your entire agent fleet to a different matrix strategy
Build your own custom matrix from scratch
See all available matrices (built-in + custom)
zen-architect, security-guardian → Opus
explorer, modular-builder, bug-hunter → Sonnet
file-ops, git-ops, session-analyst → Haiku
Foundation, superpowers, stories, design-intelligence, browser-tester, python-dev, rust-dev, LSP, containers, shadow, webruntime, and webllm agents all annotated.
One config line switches your entire routing strategy.
Data as of: March 1, 2026
Feature status: Active — deployed across 18+ repositories
Content source: Slide content authored by feature developer. Numbers reflect implementation state at time of authoring.
Key metrics:
model_role — count from agent manifests across 18+ reposamplifier-bundle-routing-matrix role definitionsamplifier-bundle-routing-matrix/data/ YAML filesBenchmark sources:
artificialanalysis.aiCopilot cost multipliers: Sourced from GitHub Copilot model pricing documentation (March 2026)
Changes from v1:
routing listGaps: Benchmark numbers are point-in-time snapshots and may shift. Copilot pricing subject to change. Agent count reflects annotated agents, not total agents in ecosystem.
See amplifier routing --help for full command reference
File issues in amplifier-bundle-routing-matrix