NIZU Agents (beta): the foundations of configurable AI assistants that can act inside your workspace, with every action recorded and attributable.
Your AI can now use several providers at once. Each kind of task can be sent to a different provider — a cheap fast one for simple steps, a stronger one for difficult ones — from Settings → AI.
Automatic failover between AI providers: if one is unavailable or rate-limited, the next one takes over instead of the task simply failing. Overnight and scheduled work no longer disappears silently.
A provider that keeps failing is temporarily set aside so it does not slow every request down, and is picked up again automatically once it recovers.
AI token usage is now recorded for every request, which is what future cost reporting per feature will be built on.
Updates
The shared AI engine now supports tools, so an assistant can look something up or carry out an action instead of only writing text. Available to every App that uses the platform AI.
Content marked as sensitive is now refused rather than quietly sent to an external AI provider. Whether that applies is governed by your GDPR setting.
Requests that are only slow or temporarily blocked are retried elsewhere, while genuine configuration mistakes (such as an invalid API key) are reported immediately instead of being retried against every provider.
Fix
DeepSeek: the default model was a name the provider no longer offers, so any workspace that had not typed a model name manually got an error on every request. It now defaults to a current model.
DeepSeek: conversations where the assistant used a tool could not be continued. They now resume correctly.
Deprecated
The private copies of the AI engine inside Marvin, WhatsAppIgnite and CMS will be retired in favour of the shared platform one. No action is needed; provider settings are unaffected.
