Workspace and global Panel share owner-private sessions, goals, messages and Turn lifecycle. Scope can be global, project or exact object. Rename, pin, export, feedback and memory APIs are available in the backend contract, while some controls are not yet exposed on all browser surfaces.
Advanexus Intelligence
AI assistance inside the platform's permission, policy and evidence boundaries.
Users describe a goal in natural language. Intelligence assembles only allowed context, selects a registered skill and lets deterministic platform services retain authority over validation and execution.
Intelligence reconstructs actor, tenant, project, permissions and RLS context on the server. The global panel receives registered page, selected-object and allow-listed hints; it does not scrape arbitrary DOM, editor content or result grids.
Seventeen versioned skills route an intent into a bounded prompt and capability set. Luna, Terra and Sol are server-owned policy profiles over the same pinned model, not user-selectable models or separate authorization domains.
A model can explain, draft and request a registered tool. The platform validates tool identity, arguments, scope and policy again. Preview or write actions require the confirmation or approval contract assigned to that exact capability.
SQL, Dataset, Analytics, Assurance, Service Desk and other registered adapters call existing application services rather than creating a parallel AI execution plane. A response-only Turn may complete with an explanation; a state-changing outcome is proven only by a completed Action or tool result linked to the canonical module.
When a user asks for an analytical Dataset or query, Intelligence uses the currently selected authorized Source, schema and metadata, states the analytical grain, dimensions, metrics and chart choice, and validates the draft before presenting or previewing it. It does not silently switch databases or treat SQL text as proof of execution.
When asynchronous mode is enabled, all profiles persist a queued Turn before execution. Durable events remain authoritative while live replay and polling provide progress. Stop, Regenerate, stale-work recovery and late-result protection retain the same bounded lifecycle.
A goal continues through owner-private state.
Workspace and global Panel share owner-private sessions, goals, messages and Turn lifecycle. Scope can be global, project or exact object. Rename, pin, export, feedback and memory APIs are available in the backend contract, while some controls are not yet exposed on all browser surfaces.
Grounded context starts with server-owned scope.
Intelligence reconstructs actor, tenant, project, permissions and RLS context on the server. The global panel receives registered page, selected-object and allow-listed hints; it does not scrape arbitrary DOM, editor content or result grids.
Actor, tenant, active Project and effective permissions
Page workflow key and exact selected-object reference
Bounded, redacted Source, SQL and canonical metadata
Available skills and tools after dependency and policy filtering
Skill and profile selection is policy, not personality theater.
Seventeen versioned skills route an intent into a bounded prompt and capability set. Luna, Terra and Sol are server-owned policy profiles over the same pinned model, not user-selectable models or separate authorization domains.
A proposal is not an action.
A model can explain, draft and request a registered tool. The platform validates tool identity, arguments, scope and policy again. Preview or write actions require the confirmation or approval contract assigned to that exact capability.
Goal → Grounded context → Proposal → Validation → Confirmation or approval → Action → Evidence
Canonical services perform the work.
SQL, Dataset, Analytics, Assurance, Service Desk and other registered adapters call existing application services rather than creating a parallel AI execution plane. A response-only Turn may complete with an explanation; a state-changing outcome is proven only by a completed Action or tool result linked to the canonical module.
Sources and SQL Console — metadata, safe SQL validation, bounded confirmed preview and SavedQuery creation
Dataset, Data Quality and Analytics — exact asset lookup, analysis guidance and narrowly registered controlled writes
ANPy — server-side AST validation of model code, never AI execution of the Python cell
Assurance and Service Desk — scoped lookup plus registered Case, package or ticket actions where the exact contract allows them
SQL assistance is grounded and dialect-aware.
When a user asks for an analytical Dataset or query, Intelligence uses the currently selected authorized Source, schema and metadata, states the analytical grain, dimensions, metrics and chart choice, and validates the draft before presenting or previewing it. It does not silently switch databases or treat SQL text as proof of execution.
Durable execution separates live progress from authority.
When asynchronous mode is enabled, all profiles persist a queued Turn before execution. Durable events remain authoritative while live replay and polling provide progress. Stop, Regenerate, stale-work recovery and late-result protection retain the same bounded lifecycle.
Capacity and usage remain explicit.
Admission control applies bounded concurrent and rolling-rate leases before provider use. Versioned global, tenant or user LIMITED and UNLIMITED policies can govern token capacity, while provider usage and versioned cost attribution remain operational records—not a hard currency budget or cross-region consensus mechanism.
The boundary is intentional.
Intelligence has no general screen scraping, web search, browser automation, autonomous workflow graph or arbitrary tool access. It does not inherit data or permissions that the current user lacks.
Drive, Sandbox and flow formalization remain response-only where no canonical executable adapter exists.
Pipeline execution, retry and cancellation, and Data Quality rule execution are not current Intelligence actions.
Analytics write is limited to a confirmed TABLE report draft over an exact DatasetVersion; it is not general chart or dashboard authoring.
Automatic documentation RAG, per-tenant provider accounts, hard currency budgets and full retention/legal-hold UI are not current capabilities.
Useful answers are concrete and honest.
A strong response uses the selected Source or object, distinguishes fact from assumption, provides validated SQL when requested, explains analytical grain and value, and stops safely when context is insufficient.
Next step
Start with the flow that cannot afford ambiguity.
Bring the systems, owners, rules, delivery obligations and evidence requirements that matter.
Discuss a critical flow