What policies really do
A policy is a rule the control plane applies to the runs of the organisation, or of a single team. The Policies screen says, for each one, what it does today — and nothing more.
What you see
- Policies by family — gates, tools, shared spaces, caps, capabilities, retention, settings —, each with its effect, written from its actual definition.
- A pill says what the control plane does with it, computed from the rule itself: Enforced, Advisory, Setting or Planned — and, for a setting, the screen where it is set.
- Every change is a new version, with who and when: the history reads back.
What each one does
| Policy | What it does | When |
|---|---|---|
| Tool approval | A call to the named tool stops at a human gate — one or two signatures, a delay, a threshold. A rule that names only a category is « Advisory »: evaluated and journaled, with no gate. | At the call |
| Human gate on runs | The run stops at its start for a human decision, where the deployment turns it on; elsewhere it is « Advisory »: the decision is recorded and the run goes. | At start |
| Cost and velocity cap | At the daily cap, the run is turned away (HTTP 429); beyond the action rate, it stops at the call; the per-run cap is checked once the run is over. After several violations, the agent is frozen until a human thaws it. | At start, at the call, after |
| Shared space | Names the agents that may share a writable space; without it, a space stays private to a single agent. | At start |
| Capability | What an agent's runs may do at all: action classes, destinations, delegation depth. | At every call and delegation |
| Retention | The daily sweep permanently deletes finished runs older than the declared period, within its scope only. | Every day |
| Organisation freeze | No run starts and no delegation leaves, until an owner thaws it. | At start |
The organisation's settings — models offered, web search, token efficiency, usage per member, escalation rules — are stored as policies too: they apply from their own screen, which the Policies screen names.
The rules of the game
- A policy can add a gate, never remove one: several rules add up, the strictest wins.
- A draft policy applies nothing.
- It targets the whole organisation, or a single team.
Who writes them
An owner, in Guardrails for the rules and in each setting's own screen. The other roles read them.
Not yet
SoonKnowledge promotion, agent behaviour, sharing — and data access or content topics outside their own screens: « Planned ». The creation drawer shows them, folded and locked, and the API refuses to turn them on.