Orchestration · internal system
The Routing Layer
Every task is classified and sent to the model that should actually do it — on capability and on cost.
Live
No public endpoint
Router
The problem it solves
Using one frontier model for everything is both wasteful and weak: you overpay for summarising a log file, and you under-think the architecture decision. Choosing manually every time does not scale, and quietly defaults to whatever is habitual.
How it works
01
Classify
what kind of task is this, really
→
02
Price
cheapest tier that can meet the quality bar
→
03
Dispatch
to a vendor CLI, an API lane or a sub-agent
→
04
Fall back
if a vendor is down, the next rung takes it
→
05
Review
high-stakes output gets a second vendor's opinion
Under the hood
- Failover and circumvention are deliberately distinguished by trigger: rerouting because a vendor is unavailable is legitimate; rerouting because a model refused on safety grounds is not.
- Bulk work is pushed down-tier aggressively; judgement, security and architecture never are.
- Escalation is one-way by default — 'not good enough' means upgrade, never downgrade.
- Every rung of the chain carries its own reasoning-effort setting rather than inheriting one global level.
Why it matters
Vendor independence is not an ideological position. It is what stops one company's pricing page from being a single point of failure in your working life.