Routing-Matrix · IncredAgents Skills

Every task routed to the right model.

A fill-in-the-blanks YAML that maps your models onto agent roles — with fallback chains, provider lanes, ban lists, and policies that force perspective diversity. Ships as a template: the structure and governance are pre-built, you fill in your models.

Ten Roles. One Sequence. Zero Guesswork.

Orchestrator, three advisors with distinct lenses, builder, QA, linter, two kaizen passes, and a logger — each with a primary model, a provider lane, and an ordered fallback chain.

Fallbacks That Walk Themselves.

Provider down, rate-limited, context window too small — the dispatcher walks the chain in order. Everything exhausted? It aborts with a list of every model tried and why each failed.

Every Provider Lane.

OpenAI, Anthropic, Google, DeepSeek, OpenRouter, local Ollama, Apple Silicon MLX, or your own custom lane — the matrix doesn't care where the model lives.

A Validator, Not A Prayer.

A script checks your filled matrix against the schema — missing fields, duplicate roles, malformed chains, banned models accidentally assigned. Green or it doesn't route.

Diversity Is Enforced, Not Hoped For.

The default policies exist because one model reviewing its own work is how blind spots ship.

No Repeats In Sequence.

No model may hold two consecutive roles — every voice in the chain sees with different training data.

Always Be Testing.

New models prove themselves in real work alongside the team — never synthetic benchmarks. The team's convergence verdict keeps them or cuts them.

Bans With Teeth.

Name a banned model in a dispatch and the dispatch gets rejected — not silently substituted with something else.

Fill In Your Models. Run The Validator.

Ships in the mod pack as a ready template with commented examples.

your machine — three commands

$ git clone https://github.com/MyAiFreedomSystems/incredagents.git

$ cd incredagents

$ python3 onboard.py

One installer wires in the whole mod pack and asks only for the keys each skill needs.