Models and runners
Your models, your runners, your rules — with the cost per change.
Work runs on your machines, through your gateway, or with your key. Routing follows your preferences inside your restrictions and records why. Every run reports what it used, against a budget that can refuse it.
Why it matters
Agent spend, tied to the change it paid for.
-
We cannot say what an agent cost us.
The invoice arrives per provider, per month. Nobody can tie it to a change, a team or an Area.
-
Confidential code goes to the default model.
A developer's tool picks the model. Nothing checks that a restricted repository stays on a local one.
-
Every vendor wants its agent in our process.
We already run several agents. We need one set of rules for all of them, not one per tool.
How it works
Route by your rules, record what was used, stop at the budget.
Routing inside your restrictions
A run takes the first of your preferences that your restrictions permit and something live can serve, and records why. Nothing is learned; the rule is yours.
- Local runners, your model gateway, or your key
- Confidential Areas kept to local venues
- One rule for runs, model calls and external agents
Usage and budgets
Usage comes from what runners and providers report, attributed to a work item, an Area, a template and a step. A budget refuses the run that would exceed it, and names itself.
- Cost is arithmetic over your price list, labelled
- Budgets per period or per work item, soft or hard
- An estimate always says it is one
External agents and standing approvals
Connect external agents over A2A, with a pinned card and skills mapped to your templates. A standing approval covers named work and ends the day a check fails.
- External agents do read work; their answers are labelled reported
- One Area, one template, limits, approvers and an expiry
- Earned over a clean window, lost on one failure
Try it
Route a run and watch the budget.
Choose an Area and a step, take the local runner offline, add runs. The run takes the first of your preferences that your restrictions permit and something live can serve, and the Area’s budget counts what it cost.
Questions about models and runners
Which agents and models can we use?
How accurate is the cost per work item?
What happens when a budget runs out?
Where do runs execute, and what leaves our machines?
Keep reading
The rest of the control plane.
Design how your company ships software.
Start with a thirty-day shadow pilot on one Area: connect, observe, read the report together, decide.