Agents and guardrails
Where agent behaviour, the guardrails and cost limits, and model routing are documented — one page per settings page, with a map from this chapter's old sections.
Concept · Updated 2026-10-17 · Describes release 0.1.164
Three pages control how the pillar agents behave in your workspace: Agent behaviour (/portal/agents), AI models (/portal/settings/models) and Guardrails & cost (/portal/guardrails). Everything on them is policy. None of it can weaken the safety floors: the write gate, PII scrubbing at the storage chokepoints and the read-only estate scan are code, not settings. Tuning changes how much the agents do and how well — never whether a human decides a write. This chapter is now one page per settings page; what counts as a write for a given connector stays in Datasources → The read/write gate in detail, and every term is defined once in the Glossary.
The pages
- Configure the agents — the seven pillar agents and the Agent behaviour page: enablement, the prompt layer, schedules, a model override and extra approval gates, with the procedures.
- Models, tiers and routing — which model serves each tier, what you inherit, the bring-your-own endpoint and its prices, smart escalation, model residency and the set-up procedure from the demo.
- Guardrails and cost limits — the write posture, the always-on floors, budgets and the cost cap, PII patterns, egress, the switches around a fix, validation checks, risk exceptions, unmanaged changes and data retention.
- Approval policy and change tiers — who may release what: the write floor, risk tiers, roles, change classes, windows and the automatic lanes.
Where the old sections went
| Old section | Now |
|---|---|
| Three layers · The pipeline · The seven pillar agents · A disabled agent still receives alerts | Configure the agents |
| Which posture your deployment runs | Guardrails and cost limits |
| The Agents page, field by field, and its three procedures | Configure the agents |
| Model tiers: which call uses which tier, what you inherit, the endpoint, prices, when the platform upgrades the model, tier suggestions | Models, tiers and routing |
| Guardrails: the floors, budgets, the cost cap, PII, egress, pre-approved operations, the switches, change safety, P1/P2 analysis, validation checks, risk exceptions, unmanaged changes, data retention | Guardrails and cost limits |
| Where the money went: cached prompt reads | Models, tiers and routing |
| Model residency policy | Models, tiers and routing |
| Use cases, troubleshooting, limits and gotchas | Each page keeps the rows about its own settings page |