# Settings & tiered configuration

> Operational limits and defaults resolve through tiers — cluster, namespace, agent — where a cluster ceiling can clamp what a class is allowed to loosen.

Source: https://openap.org/docs/settings-config

Not every knob belongs on every AgentClass. Budgets, tool-guard limits, content inspectors, and the like are
 **settings** that resolve through tiers — a cluster operator sets defaults and hard ceilings,
and an individual agent narrows within them. It's how one team can hand out agents without handing out the
ability to remove their own guardrails.

## Defaults flow down; ceilings clamp up

Settings resolve from the **cluster**, through the **namespace**, down to the **AgentClass** — the most specific
tier that sets a value wins. But there are two kinds of setting, and they compose differently:

- A **default** is a starting value a lower tier may override. Set a default token budget cluster-wide, and an
  agent can raise or lower it.
- A **ceiling** is a hard limit a lower tier may *not* exceed. Set a ceiling, and an agent can go stricter but
  never looser — it clamps whatever the class tries to loosen.

So a platform team pins the outer bounds once (the ceilings), ships sensible defaults, and lets each agent tune
inside that envelope.

## What's tiered

The tiered surface is the operational policy that shouldn't live per-agent: [budgets](https://openap.org/docs/session-lifecycle),
[tool-guard breakers, rate limits, and data budgets](https://openap.org/docs/budgets-breakers), [content inspectors](https://openap.org/docs/content-guards-egress),
the default model, and requirements like "subagent digest pins are mandatory." An agent reads its **effective**
settings — the resolved result of all tiers — so what actually applies is a single computed answer, not
something you reconstruct by hand.

## Setting them

`oap settings wizard` walks you through the security-relevant defaults and ceilings interactively; `oap settings
apply` applies a settings file. The result is recorded so you can see the effective policy an agent is running
under — including in the [admin console](https://openap.org/docs/admin-dashboard), which shows the resolved value per setting:

![Config › Settings — the tiered defaults and ceilings, resolved and shown per setting.](https://openap.org/media/admin-settings.png)

On the desktop app the same cluster settings are editable directly from the menubar's **Settings…** window. Its
**Cluster** tab groups the model catalog, limits, and defaults into sub-tabs, validates a change against the live
cluster before you save, and keeps the save controls pinned in reach as the forms grow:

![Desktop → Settings → Cluster — the model catalog, limits and defaults as sub-tabs, validated against the live cluster before you save.](https://openap.org/media/desktop-settings.png)

> **Security: An agent can't widen its own guardrails**
>
> The ceiling/default split is the point: safety limits set above an agent are ones it can only tighten. A
> compromised or over-eager class can't raise its token budget past the cluster cap or turn off a required
> content inspector — the resolver clamps it, and the clamp is enforced, not advisory.
