How it works

Look up the model fresh before every run

An alias is a row an operator edits on the admin screen, and the next run picks up the change with no deploy. A single run can also override its own selection without touching the standing alias for everyone else.

  • One lookup, every Agent Continuous Improvement Coach reasoning Investigation Engine reasoning Demand Planner default
  • Alias updated 09:14

Model lookup

Keep the model choice in one place

No Agent carries a model name inside it. Each one asks the Model Router at run time and gets an answer back, or a safe default if nothing has been set yet. One lookup covers every Agent and every Module that uses a model.

  • No model is named inside an Agent, anywhere
  • One lookup covers every Agent and every AI-enabled Module
  • A bootstrap default answers when no alias is set yet
  • R. Vantasselrepointed alias 09:40
  • Picked up by RUN-30150

Model alias

Change the model without a release

Edit the alias on the admin screen and the next run picks it up, with no deploy. The prompt resolves alongside the model, so it comes back as one answer.

  • Aliases are rows an operator edits on the admin screen
  • The prompt resolves alongside the model, as one answer
  • The change takes effect on the next run, with no deploy
  • RUN-30149: Quality ops Emberline R2 Standard Override
  • Alias: qa-engine default

Router monitoring

Override one run, keep the rest

A single run can name its own selection without touching the standing alias. If that selection isn’t available, the call fails over automatically and the run keeps going. The failover is recorded, and everyone else’s alias stays as it was.

  • A per-run override applies to that run only
  • Failover moves the call and records the switch
  • The standing alias is untouched either way
  • Downgrades, today15

Budget guard

Downgrade on budget instead of failing

A budget guard checks spend at the moment the selection is made. Go over the line and the run continues on a cheaper selection instead of stopping. Every downgrade is written down with the run it affected.

  • Spend is checked at the moment the selection is made
  • Over budget, the run downgrades rather than fails
  • Each downgrade is written with the run it affected
  • Decision, RUN-30149reason: budget downgrade

Routing diagnostics

Trace any run back to its reason

Each selection writes a machine-readable reason alongside it, so a past run traces back to what it used and why. An alias, an override, a failover and a budget downgrade each read differently in diagnostics, so you always know which one fired.

  • The reason is a fixed code a report can read
  • Diagnostics show the selection and reason for any run
  • Alias, override, failover, and budget each read differently
  • Alias, by Agent Quality (Auditor) qa-engine default Emberline R2

Per-Agent alias

Pick the model for each Agent by its alias

The Model Router selects by the alias set for that Agent. It doesn’t read a request and decide the work needs a bigger model. Different behaviour only comes from setting a different alias.

  • No routing by task complexity or difficulty tier
  • Selection follows the alias an operator set
  • Different behaviour comes from a different alias

What an admin sees when a run needs a different model

    • Run overrides, today RUN-30149 Standard Override RUN-30162 Standard Alias RUN-30170 Emberline R2 Failover
    • Alias untouched

    One run picks its own model, nothing else changes.

    A single run can name its own model without touching the standing alias. If that choice is not available, the call fails over and keeps going, and the switch gets written down.

    See override and failover
    • Downgrades today 6
    • Decision, RUN-30170reason: budget downgrade

    Every downgrade is written down with its run.

    A budget guard checks spend the moment a model gets picked. Go over the line and the run continues on a cheaper model instead of stopping, and the reason is recorded alongside it.

    See how downgrades log

FAQ

Ask which model your Agents actually use.

30 minutes on your Agents and how the Model Router records and traces each choice.