diff --git a/app/components/assistant/message-body.tsx b/app/components/assistant/message-body.tsx
index 8e9540c..5003978 100644
--- a/app/components/assistant/message-body.tsx
+++ b/app/components/assistant/message-body.tsx
@@ -1,30 +1,53 @@
-// Renders an assistant message: markdown for prose, and a small "Ran N
-// actions" pill in place of any ```action``` fenced blocks (which are the
-// machine-readable instructions the bus has already executed).
+// Renders an assistant message: markdown for prose, plus pills for any
+// command-bus action blocks or native tool calls attached to the message.
+// Tool-role messages (results from a function call) render as a JSON card.
 
 import { useMemo } from "react"
 import ReactMarkdown from "react-markdown"
-import { Sparkles } from "lucide-react"
+import { Sparkles, Wrench } from "lucide-react"
 
 import { extractActionBlocks } from "@crema/action-bus"
+import { stripToolCallTags, type ToolCall } from "@crema/llm-ui"
 
 const ACTION_BLOCK_RE = /```action\s*\n[\s\S]*?```/g
 
-export function MessageBody({ content }: { content: string }) {
+export type MessageBodyProps = {
+  content: string
+  /** When set, render as a tool-result card. */
+  isToolResult?: boolean
+  /** Native tool calls attached to this assistant message, if any. */
+  toolCalls?: ToolCall[]
+}
+
+export function MessageBody({ content, isToolResult, toolCalls }: MessageBodyProps) {
   const { prose, actionCount } = useMemo(() => {
     const blocks = extractActionBlocks(content)
+    const cleaned = stripToolCallTags(content).replace(ACTION_BLOCK_RE, "").trim()
     return {
-      prose: content.replace(ACTION_BLOCK_RE, "").trim(),
+      prose: cleaned,
       actionCount: blocks.length,
     }
   }, [content])
 
+  if (isToolResult) {
+    return (
+      <div className="rounded-md border border-border/60 bg-muted/40 px-3 py-2 text-xs">
+        <div className="mb-1 flex items-center gap-1.5 font-medium text-muted-foreground">
+          <Wrench className="size-3" />
+          Tool result
+        </div>
+        <pre className="overflow-x-auto whitespace-pre-wrap font-mono text-[11px] leading-snug text-foreground/80">
+          {content}
+        </pre>
+      </div>
+    )
+  }
+
   return (
     <div className="prose prose-sm max-w-none dark:prose-invert">
       {prose && (
         <ReactMarkdown
           components={{
-            // Tight overrides — keep paragraphs compact in chat bubbles
             p: ({ children }) => <p className="my-1.5 leading-relaxed">{children}</p>,
             code: ({ children, className }) => {
               const isBlock = className?.startsWith("language-")
@@ -63,6 +86,15 @@ export function MessageBody({ content }: { content: string }) {
           Ran {actionCount} action{actionCount > 1 ? "s" : ""}
         </span>
       )}
+      {toolCalls && toolCalls.length > 0 && (
+        <span
+          className="mt-1 ml-1 inline-flex items-center gap-1.5 rounded-full border border-amber-400/40 bg-amber-400/10 px-2.5 py-0.5 text-xs font-medium text-amber-700 dark:border-amber-300/60 dark:bg-amber-300/10 dark:text-amber-200"
+          title="Tool call dispatched"
+        >
+          <Wrench className="size-3" />
+          Called {toolCalls.map((c) => c.name).join(", ")}
+        </span>
+      )}
     </div>
   )
 }
diff --git a/app/lib/admin-context.ts b/app/lib/admin-context.ts
new file mode 100644
index 0000000..6489cba
--- /dev/null
+++ b/app/lib/admin-context.ts
@@ -0,0 +1,76 @@
+// Shared state surface that any admin page can publish to so the assistant
+// can read live data without scraping the DOM.
+//
+// Pages call `useRegisterAdminContext("tenants", { tenants: [...] })` while
+// mounted; the assistant calls `getAdminContextSnapshot()` each turn to
+// inject a structured snapshot into the system prompt.
+
+import { useEffect } from "react"
+
+type Surface = Record<string, unknown>
+
+export type AdminContextSnapshot = {
+  route: string
+  surfaces: Record<string, Surface>
+}
+
+const surfaces = new Map<string, Surface>()
+
+export function publishAdminSurface(name: string, data: Surface): void {
+  surfaces.set(name, data)
+  if (typeof window !== "undefined") {
+    ;(window as unknown as { __adminContext?: unknown }).__adminContext = getAdminContextSnapshot()
+  }
+}
+
+export function clearAdminSurface(name: string): void {
+  surfaces.delete(name)
+  if (typeof window !== "undefined") {
+    ;(window as unknown as { __adminContext?: unknown }).__adminContext = getAdminContextSnapshot()
+  }
+}
+
+export function getAdminContextSnapshot(): AdminContextSnapshot {
+  const route = typeof window !== "undefined" ? window.location.pathname : ""
+  return {
+    route,
+    surfaces: Object.fromEntries(surfaces.entries()),
+  }
+}
+
+/**
+ * Render a snapshot as a markdown block for the LLM system prompt.
+ * Keeps it compact: route, then one section per surface with JSON.
+ */
+export function formatAdminContextForPrompt(snapshot = getAdminContextSnapshot()): string {
+  const sections: string[] = [`Admin context (read-only — for answering factual questions):`]
+  sections.push(`Route: ${snapshot.route || "?"}`)
+  const names = Object.keys(snapshot.surfaces)
+  if (names.length === 0) {
+    sections.push(`Surfaces: (none registered)`)
+  } else {
+    for (const name of names) {
+      const json = safeJson(snapshot.surfaces[name])
+      sections.push(`Surface "${name}":\n${json}`)
+    }
+  }
+  return sections.join("\n\n")
+}
+
+function safeJson(value: unknown): string {
+  try {
+    const text = JSON.stringify(value, null, 2)
+    if (text.length > 4000) return text.slice(0, 4000) + "\n…(truncated)"
+    return text
+  } catch {
+    return "(unserializable)"
+  }
+}
+
+/** Hook: publish a surface while the component is mounted. */
+export function useRegisterAdminContext(name: string, data: Surface): void {
+  useEffect(() => {
+    publishAdminSurface(name, data)
+    return () => clearAdminSurface(name)
+  }, [name, data])
+}
diff --git a/app/lib/admin-tools.ts b/app/lib/admin-tools.ts
new file mode 100644
index 0000000..d19cec3
--- /dev/null
+++ b/app/lib/admin-tools.ts
@@ -0,0 +1,161 @@
+// Curated tool surface the assistant can call. The LLM emits a fenced
+// ```tool block with one JSON object per line; we parse, execute via
+// arcadia-client, and feed results back as the next user turn.
+//
+// Each tool is a named function with documented args. The LLM never sees
+// raw HTTP — only the menu below.
+
+import type { ArcadiaClient } from "@crema/arcadia-client"
+import type { Tool, ToolCall as LLMToolCall } from "@crema/llm-ui"
+
+import { getTenant, listTenants, type Tenant } from "~/lib/arcadia/tenants"
+
+export type ToolCall = {
+  name: string
+  args: Record<string, unknown>
+}
+
+export type ToolResult = {
+  name: string
+  args: Record<string, unknown>
+  ok: boolean
+  data?: unknown
+  error?: string
+}
+
+type ToolDef = {
+  name: string
+  description: string
+  parameters: Record<string, unknown> // JSON Schema for OpenAI tool calling
+  isWrite: boolean
+  run: (args: Record<string, unknown>, ctx: ToolCtx) => Promise<unknown>
+}
+
+type ToolCtx = { arcadia: ArcadiaClient }
+
+const TOOLS: ToolDef[] = [
+  {
+    name: "list_tenants",
+    description:
+      "List every tenant on this arcadia deployment. Returns id, slug, name, status, plan, inserted_at. Call this for any question about tenant counts, statuses, or which tenants exist.",
+    parameters: {
+      type: "object",
+      properties: {},
+      additionalProperties: false,
+    },
+    isWrite: false,
+    run: async (_args, { arcadia }) => {
+      const tenants = await listTenants(arcadia)
+      return tenants.map(summarize)
+    },
+  },
+  {
+    name: "get_tenant",
+    description:
+      "Fetch a single tenant by slug (preferred) or id. Returns the tenant summary or null if not found.",
+    parameters: {
+      type: "object",
+      properties: {
+        slug: { type: "string", description: "The tenant's slug (e.g. 'acme', 'platform-admin')." },
+        id: { type: "string", description: "The tenant's UUID. Use only when the slug is unknown." },
+      },
+      additionalProperties: false,
+    },
+    isWrite: false,
+    run: async (args, { arcadia }) => {
+      const slug = typeof args.slug === "string" ? args.slug : null
+      const id = typeof args.id === "string" ? args.id : null
+      if (!slug && !id) throw new Error("get_tenant requires { slug } or { id }")
+      if (id) {
+        try {
+          return summarize(await getTenant(arcadia, id))
+        } catch {
+          return null
+        }
+      }
+      const tenants = await listTenants(arcadia)
+      const found = tenants.find((t) => t.slug === slug)
+      return found ? summarize(found) : null
+    },
+  },
+]
+
+/** OpenAI-format tool list to pass into ChatRequest.tools. */
+export function getOpenAITools(): Tool[] {
+  return TOOLS.map((t) => ({
+    name: t.name,
+    description: t.description,
+    parameters: t.parameters,
+  }))
+}
+
+function summarize(t: Tenant) {
+  return {
+    id: t.id,
+    slug: t.slug,
+    name: t.name,
+    status: t.status,
+    plan: t.plan?.name ?? null,
+    inserted_at: t.inserted_at,
+  }
+}
+
+const TOOL_BY_NAME = new Map(TOOLS.map((t) => [t.name, t]))
+
+function safeJson(value: unknown): string {
+  try {
+    const text = JSON.stringify(value, null, 2)
+    if (text.length > 6000) return text.slice(0, 6000) + "\n…(truncated)"
+    return text
+  } catch {
+    return "(unserializable)"
+  }
+}
+
+/** Run a list of provider-native tool calls and return `tool` role messages
+ *  ready to push back into useChat history. */
+export async function runLLMToolCalls(
+  calls: LLMToolCall[],
+  ctx: ToolCtx,
+  opts: { allowWrites?: boolean } = {},
+): Promise<{
+  results: ToolResult[]
+  toolMessages: { role: "tool"; content: string; toolCallId: string; name: string }[]
+}> {
+  const results: ToolResult[] = []
+  const toolMessages: { role: "tool"; content: string; toolCallId: string; name: string }[] = []
+  for (const call of calls) {
+    const def = TOOL_BY_NAME.get(call.name)
+    let parsed: Record<string, unknown> = {}
+    try {
+      parsed = call.arguments ? (JSON.parse(call.arguments) as Record<string, unknown>) : {}
+    } catch {
+      const err = `Could not parse arguments JSON: ${call.arguments}`
+      results.push({ name: call.name, args: {}, ok: false, error: err })
+      toolMessages.push({ role: "tool", content: JSON.stringify({ error: err }), toolCallId: call.id, name: call.name })
+      continue
+    }
+    if (!def) {
+      const err = `Unknown tool: ${call.name}`
+      results.push({ name: call.name, args: parsed, ok: false, error: err })
+      toolMessages.push({ role: "tool", content: JSON.stringify({ error: err }), toolCallId: call.id, name: call.name })
+      continue
+    }
+    if (def.isWrite && !opts.allowWrites) {
+      const err = "Write tools require user confirmation."
+      results.push({ name: call.name, args: parsed, ok: false, error: err })
+      toolMessages.push({ role: "tool", content: JSON.stringify({ error: err }), toolCallId: call.id, name: call.name })
+      continue
+    }
+    try {
+      const data = await def.run(parsed, ctx)
+      results.push({ name: call.name, args: parsed, ok: true, data })
+      toolMessages.push({ role: "tool", content: safeJson(data), toolCallId: call.id, name: call.name })
+    } catch (err) {
+      const msg = err instanceof Error ? err.message : String(err)
+      results.push({ name: call.name, args: parsed, ok: false, error: msg })
+      toolMessages.push({ role: "tool", content: JSON.stringify({ error: msg }), toolCallId: call.id, name: call.name })
+    }
+  }
+  return { results, toolMessages }
+}
diff --git a/app/lib/agents.ts b/app/lib/agents.ts
index 9a058f0..0ff125d 100644
--- a/app/lib/agents.ts
+++ b/app/lib/agents.ts
@@ -13,39 +13,39 @@ export type Agent = {
 
 export const DEFAULT_AGENTS: Agent[] = [
   {
-    id: "generalist",
+    id: "operator",
     name: "Atlas",
-    role: "Generalist",
+    role: "Platform Operator",
     prompt:
-      "You handle anything: chat, planning, summaries, casual questions. Match the user's tone. Keep replies as long as the task deserves — terse for quick questions, detailed when explaining.",
+      "You're the platform admin's day-to-day operator inside Arcadia Admin. Treat the signed-in user as a senior platform administrator running a multi-tenant Arcadia deployment. Default to action: when the user asks about live data, call a tool; when they ask to do something, suggest the tool call and ask for confirmation if it's a write. Prefer tenant slugs over UUIDs in conversation. Keep replies tight — operators read fast.",
   },
   {
-    id: "coder",
-    name: "Forge",
-    role: "Software engineer",
+    id: "auditor",
+    name: "Ledger",
+    role: "Auditor",
     prompt:
-      "You are a senior software engineer. Write idiomatic, well-typed code. Prefer concrete examples over abstract advice. When asked to fix a bug, identify root cause before patching. Use markdown code blocks with language tags. Mention edge cases briefly when relevant.",
+      "You're an audit-focused assistant inside Arcadia Admin. Specialise in audit logs, access reviews, and 'who did what when' questions. Always cite the actor_type (user / platform_admin / api_key / system) and timestamp when summarising audit entries. Be cautious about claims you can't back with a tool result — call a tool first.",
   },
   {
-    id: "writer",
-    name: "Inkwell",
-    role: "Writer",
+    id: "triage",
+    name: "Beacon",
+    role: "Incident Triage",
     prompt:
-      "You are a prose writer. Produce vivid, well-paced text — short stories, copy, emails, essays. Vary sentence length. Show, don't tell. When the user asks for a draft, deliver the draft, not a description of it.",
+      "You're an incident-triage assistant inside Arcadia Admin. When the user reports a problem (a tenant member can't sign in, a billing call is 402'ing, a webhook is failing), walk the diagnostic tree: identify the tenant, check tenant status, check the user's roles, check the billing-config / api-metering / feature-flag overrides as relevant. Suggest impersonation only when it's the right escalation. Keep a clear hypothesis → check → result rhythm.",
   },
   {
-    id: "researcher",
-    name: "Pilot",
-    role: "Researcher",
+    id: "analyst",
+    name: "Tally",
+    role: "Platform Analyst",
     prompt:
-      "You are a careful researcher. Structure answers as: claim → evidence → caveat. Distinguish what is well-established from what is uncertain. Refuse to fabricate citations — if you don't know, say so.",
+      "You're an analyst inside Arcadia Admin. Answer numerical and aggregate questions across the platform: tenant counts by status, plan distribution, audit-log volume, growth. Always pull live data via tools — never guess from stale snapshots. Present findings in plain prose first, then a small table when the breakdown helps.",
   },
   {
     id: "ui-driver",
     name: "Cursor",
     role: "UI Operator",
     prompt:
-      "You specialize in driving this app's UI on the user's behalf. Prefer doing over explaining. When the user asks for an action, emit an action block immediately. When they ask a question about the app, answer concisely and offer to do it.",
+      "You specialise in driving Arcadia Admin's UI on the operator's behalf. Prefer doing over explaining. When the user asks for an action that maps to a UI element, emit an action block immediately (using `data-action` ids the host has documented). For data questions, prefer tool calls over UI navigation.",
   },
 ]
 
@@ -64,6 +64,14 @@ function isAgent(v: unknown): v is Agent {
   )
 }
 
+// Old Vibespace agent ids — used to auto-migrate operators stuck on the
+// generic defaults from before Arcadia Admin had its own personas.
+const LEGACY_AGENT_IDS = new Set(["generalist", "coder", "writer", "researcher"])
+
+function isLegacyDefaultSet(agents: Agent[]): boolean {
+  return agents.some((a) => LEGACY_AGENT_IDS.has(a.id))
+}
+
 function readFromStorage(): Agent[] {
   if (typeof window === "undefined") return DEFAULT_AGENTS
   try {
@@ -72,7 +80,14 @@ function readFromStorage(): Agent[] {
     const parsed = JSON.parse(raw)
     if (!Array.isArray(parsed)) return DEFAULT_AGENTS
     const cleaned = parsed.filter(isAgent)
-    return cleaned.length > 0 ? cleaned : DEFAULT_AGENTS
+    if (cleaned.length === 0) return DEFAULT_AGENTS
+    if (isLegacyDefaultSet(cleaned)) {
+      // Auto-migrate: stored set still contains pre-arcadia personas.
+      localStorage.setItem(STORAGE_KEY, JSON.stringify(DEFAULT_AGENTS))
+      localStorage.removeItem(ACTIVE_KEY)
+      return DEFAULT_AGENTS
+    }
+    return cleaned
   } catch {
     return DEFAULT_AGENTS
   }
diff --git a/app/lib/arcadia-knowledge.ts b/app/lib/arcadia-knowledge.ts
new file mode 100644
index 0000000..59abf23
--- /dev/null
+++ b/app/lib/arcadia-knowledge.ts
@@ -0,0 +1,35 @@
+// Domain primer baked into the assistant's system prompt so it understands
+// what arcadia-app is, what platform admins do, and how the data model fits
+// together. Keep this tight — it costs context tokens on every turn.
+
+export const ARCADIA_KNOWLEDGE = `Arcadia (the backend you administer):
+
+Arcadia is a multi-tenant SaaS backend (Elixir/Phoenix umbrella, OpenAPI at /api/v1, server-rendered platform UI at /platform/*). This admin app (Arcadia Admin) is one of several clients — it talks to Arcadia over JSON, scoped by an X-Tenant-ID header and a Bearer JWT.
+
+Core entities and how they relate:
+
+- **Tenant** — an isolated workspace (a customer org). Identified by a slug (e.g. "acme", "platform-admin", "default") and a UUID id. Owns its own users, roles, billing config, branding, settings. Most data is tenant-scoped.
+- **Platform admin** — a separate identity that lives in the platform_admins table, NOT in any tenant. The signed-in operator using this app is one. Can read/write across all tenants. The first one is bootstrapped via /setup; \`is_root: true\` flags the original.
+- **User** — a member of a single tenant. Has email + password (or SSO), system roles (\`admin\` / \`user\` / \`viewer\`) plus optional custom roles. Login goes through POST /api/v1/auth/login with the tenant slug in X-Tenant-ID.
+- **Role** — permission bundle scoped to a tenant. \`admin\` / \`user\` / \`viewer\` are seeded as system roles per tenant. Permissions are wildcard-ish strings (e.g. \`tenants:read\`, \`*\`).
+- **Plan** — subscription tier attached to a tenant: name + limits (seats, storage, API quota). Drives billing.
+- **Audit log entry** — append-only record of who did what. \`actor_type\` is one of: \`user\`, \`platform_admin\`, \`api_key\`, \`system\`. Per-tenant and platform-wide entries coexist.
+- **Feature flag** — boolean / variant gate. Platform-wide default + per-tenant override.
+- **Storage / billing config / SSO IdP / inbound webhook / API quota / data retention policy / approval workflow / announcement** — per-tenant or platform-level configurations the operator can manage.
+
+Tenant lifecycle (status field):
+
+- **active** — normal operation. Members can sign in. Default state.
+- **suspended** — members blocked from signing in. Reversible: activate to restore. Use for temporary holds (overdue invoice, abuse investigation).
+- **deactivated** — stronger stop. Treat as effectively closed; usually flagged as terminal even if technically reversible. Use only when offboarding.
+
+Things to keep in mind when assisting:
+
+- Prefer tenant **slugs** in user-facing language ("the acme tenant"); slugs are stable, ids are UUIDs that aren't useful to humans.
+- "Platform admin" ≠ "admin role inside a tenant". The first acts cross-tenant; the second is scoped to one tenant.
+- Writes are auditable. Suggest the user double-check tenant slug and impact before suspend/deactivate. Deactivate is harsher than suspend — only use when clearly intended.
+- The operator can impersonate tenant users for debugging (POST /api/v1/admin/impersonate/:user_id) — surface this when they ask "why can't user X log in".
+- Quotas / rate cards / billing config errors usually surface as 402/403 from /api/v1 endpoints — diagnose by checking the tenant's billing-config and api-metering quotas.
+- The reference Phoenix app lives at \`reference/arcadia-app/\` in the workspace; its OpenAPI spec is at /api/openapi (sync via \`node ../lib-arcadia-client/scripts/sync-spec.mjs\`).
+
+When the user asks something that maps to a tool, call it. When they ask about a concept, explain it from this primer in plain language. When they ask to do something destructive, summarise the impact in one sentence and ask for confirmation before suggesting a tool call.`
diff --git a/app/lib/threads.ts b/app/lib/threads.ts
index 9a44111..782099d 100644
--- a/app/lib/threads.ts
+++ b/app/lib/threads.ts
@@ -4,10 +4,16 @@
 import { useEffect, useSyncExternalStore } from "react"
 
 export type ThreadMessage = {
-  role: "user" | "assistant"
+  role: "user" | "assistant" | "tool"
   content: string
   /** Persona that authored this assistant message (omitted for user msgs). */
   agentId?: string
+  /** Native tool calls attached to an assistant message. */
+  toolCalls?: { id: string; name: string; arguments: string }[]
+  /** Tool role only — id of the matching assistant tool_call. */
+  toolCallId?: string
+  /** Tool role only — function name. */
+  name?: string
 }
 
 export type Thread = {
diff --git a/app/routes/ai.tsx b/app/routes/ai.tsx
index 8217e0c..4a2c240 100644
--- a/app/routes/ai.tsx
+++ b/app/routes/ai.tsx
@@ -53,7 +53,6 @@ import {
 } from "~/components/ui/popover"
 import { useLLMSettings } from "~/lib/llm-settings"
 import {
-  composeSystemPrompt,
   loadActiveAgentId,
   saveActiveAgentId,
   useAgents,
@@ -62,6 +61,11 @@ import {
 import { addLibraryItem } from "~/lib/library"
 import { Avatar, AvatarFallback } from "~/components/ui/avatar"
 import { pageTitle } from "~/lib/page-meta"
+import { useArcadiaClient } from "@crema/arcadia-client"
+import type { ToolCall } from "@crema/llm-ui"
+import { getOpenAITools, runLLMToolCalls } from "~/lib/admin-tools"
+import { ARCADIA_KNOWLEDGE } from "~/lib/arcadia-knowledge"
+import { formatAdminContextForPrompt } from "~/lib/admin-context"
 
 const SNAPSHOT_KEY = "crema.ai.snapshot"
 type StoredMessage = { role: "user" | "assistant"; content: string }
@@ -221,12 +225,52 @@ function ChatSurface({
   isMock: boolean
   onRetryProbe: () => void
 }) {
-  const baseSystem =
-    "You are a helpful AI assistant. Be concise, warm, and direct."
-  const systemPrompt = composeSystemPrompt(baseSystem, activeAgent)
-  const { messages, setMessages, send, abort, isStreaming, reset } = useChat({
+  const persona = activeAgent
+    ? `Active persona: ${activeAgent.name} — ${activeAgent.role}\n${activeAgent.prompt}`
+    : ""
+  const systemPrompt = [
+    "You are the operator's assistant inside Arcadia Admin. Be precise and direct. You have native function tools attached to this conversation — call them whenever the user asks about live platform state (counts, statuses, listings, lookups). Never invent tenant slugs, user counts, or statuses; if you need data, call a tool.",
+    ARCADIA_KNOWLEDGE,
+    persona,
+    formatAdminContextForPrompt(),
+  ]
+    .filter(Boolean)
+    .join("\n\n")
+  const arcadia = useArcadiaClient()
+  const { messages, setMessages, send, continueChat, abort, isStreaming, reset } = useChat({
     system: systemPrompt,
   })
+
+  // Auto tool-loop using native function calls.
+  const toolIterationsRef = useRef(0)
+  const processedTurnRef = useRef(-1)
+  const prevStreamingRef = useRef(isStreaming)
+  const MAX_TOOL_ITERATIONS = 3
+  useEffect(() => {
+    const justFinished = prevStreamingRef.current && !isStreaming
+    prevStreamingRef.current = isStreaming
+    if (!justFinished) return
+    const lastIdx = messages.length - 1
+    if (lastIdx < 0) return
+    const last = messages[lastIdx]
+    if (last.role !== "assistant") return
+    if (processedTurnRef.current === lastIdx) return
+    processedTurnRef.current = lastIdx
+    const calls = last.toolCalls ?? []
+    if (calls.length === 0) {
+      toolIterationsRef.current = 0
+      return
+    }
+    if (toolIterationsRef.current >= MAX_TOOL_ITERATIONS) return
+    toolIterationsRef.current += 1
+    void (async () => {
+      const { toolMessages } = await runLLMToolCalls(calls, { arcadia })
+      void continueChat(toolMessages, {
+        system: systemPrompt,
+        tools: getOpenAITools(),
+      })
+    })()
+  }, [messages, isStreaming, arcadia, continueChat, systemPrompt])
   const { complete: completeOneShot, isLoading: compacting } = useCompletion()
   const [input, setInput] = useState("")
   const [showPromptOpen, setShowPromptOpen] = useState(false)
@@ -304,12 +348,12 @@ function ChatSurface({
     const text = messages[lastUserIdx].content
     setMessages(messages.slice(0, lastUserIdx))
     // Defer so the state flush completes before send() reads `messages`.
-    setTimeout(() => void send(text), 0)
+    setTimeout(() => void send(text, { tools: getOpenAITools() }), 0)
   }, [messages, setMessages, send, isStreaming])
 
   const continueLast = useCallback(() => {
     if (isStreaming || messages.length === 0) return
-    void send("Please continue your previous reply.")
+    void send("Please continue your previous reply.", { tools: getOpenAITools() })
   }, [isStreaming, messages.length, send])
 
   const compactConversation = useCallback(async () => {
@@ -396,7 +440,7 @@ function ChatSurface({
     if (!text || isStreaming) return
     setInput("")
     stickRef.current = true
-    void send(text)
+    void send(text, { tools: getOpenAITools() })
   }, [input, isStreaming, send])
 
   const isEmpty = messages.length === 0
@@ -426,9 +470,16 @@ function ChatSurface({
       ) : (
         <div className="flex-1 px-4 py-6 sm:px-6">
           <div className="mx-auto flex w-full max-w-3xl flex-col gap-6">
-            {messages.map((m, i) => (
-              <MessageRow key={i} role={m.role} content={m.content} />
-            ))}
+            {messages
+              .filter((m) => m.role !== "system")
+              .map((m, i) => (
+                <MessageRow
+                  key={i}
+                  role={m.role as "user" | "assistant" | "tool"}
+                  content={m.content}
+                  toolCalls={m.toolCalls}
+                />
+              ))}
             {isStreaming && messages.at(-1)?.role !== "assistant" && (
               <div className="self-start">
                 <TypingIndicator />
@@ -499,10 +550,19 @@ function ChatSurface({
 function MessageRow({
   role,
   content,
+  toolCalls,
 }: {
-  role: "user" | "assistant"
+  role: "user" | "assistant" | "tool"
   content: string
+  toolCalls?: ToolCall[]
 }) {
+  if (role === "tool") {
+    return (
+      <div className="self-start max-w-[80ch]">
+        <MessageBody content={content} isToolResult />
+      </div>
+    )
+  }
   if (role === "user") {
     return (
       <div className="self-end">
@@ -520,7 +580,7 @@ function MessageRow({
   }
   return (
     <div className="self-start max-w-[80ch]">
-      <MessageBody content={content} />
+      <MessageBody content={content} toolCalls={toolCalls} />
     </div>
   )
 }
diff --git a/app/routes/assistant.tsx b/app/routes/assistant.tsx
index 695db76..606562d 100644
--- a/app/routes/assistant.tsx
+++ b/app/routes/assistant.tsx
@@ -31,7 +31,9 @@ const PROBE_TIMEOUT_MS = 3000
 // "Available actions" in the system prompt only lists what's on screen NOW;
 // this catalog tells the model what exists elsewhere so it can plan
 // multi-step flows (navigate → wait_for → fill → click) in a single block.
-const UI_CONTROL_PREFACE = `You are the assistant inside this app and you can drive the UI. A virtual cursor is shown to the user — every step should target an element so the cursor visibly moves.
+const UI_CONTROL_PREFACE = `You are the operator's assistant inside Arcadia Admin — the platform-admin web app for the Arcadia multi-tenant SaaS backend (Phoenix, /api/v1). The signed-in user is a platform administrator. Help them inspect and manage tenants, users, billing, audit logs, feature flags, and other platform surfaces.
+
+You can both answer factual questions about the current state (use the "Admin context" block below) and drive the UI. A virtual cursor is shown to the user — every step should target an element so the cursor visibly moves.
 
 Rules:
 - Prefer \`click nav-<page>\` (e.g. \`click nav-settings\`) over \`navigate <path>\` so the user sees the cursor travel to the sidebar.
@@ -42,7 +44,7 @@ Rules:
 
 Known action ids across the app (use these even if not in "Available actions" — the page may not be mounted yet):
 
-Sidebar / nav: nav-overview, nav-resources, nav-activity, nav-assistant, nav-library, nav-settings, sidebar-toggle, mobile-nav-toggle
+Sidebar / nav: nav-overview, nav-tenants, nav-resources, nav-activity, nav-assistant, nav-library, nav-settings, sidebar-toggle, mobile-nav-toggle
 Appbar: appbar-search (input), appbar-scripts, appbar-font-size, appbar-surface, appbar-background, theme-toggle, appbar-notifications, appbar-avatar
 Account menu (after click appbar-avatar): avatar-profile (→ /profile), avatar-settings, avatar-help, avatar-signout
 Profile page: profile-avatar-upload, profile-avatar-remove, profile-name, profile-email, profile-title, profile-bio, profile-signature, profile-default-agent, profile-save, profile-revert, profile-reset
@@ -54,6 +56,7 @@ Assistant agent picker: assistant-agent (dropdown — click to switch persona)
 Assistant page: assistant-model, assistant-agent, assistant-thread, assistant-thread-new, assistant-thread-switch-<id>, assistant-thread-rename-<id>, assistant-thread-delete-<id>, assistant-ui-control, assistant-compact, assistant-restore-compact, assistant-regenerate, assistant-continue, assistant-show-prompt, assistant-copy-md, assistant-export-md, assistant-save-library, assistant-compare, assistant-handoff-<id>, assistant-stop, assistant-clear, assistant-retry-probe, assistant-actions (open the kebab popover first to reveal the rest), assistant-voice (mic), assistant-msg-pin-<i>, assistant-msg-edit-<i>, assistant-msg-speak-<i>
 Library page: library-search, library-open-<id>, library-copy-<id>, library-download-<id>, library-delete-<id>
 Resources page: resources-search, resources-new-name, resources-create, resources-status-<id>, resources-delete-<id>
+Tenants page: tenants-refresh, tenants-search (input), tenants-create. Per-row (use the tenant's slug — see the "tenants" surface in Admin context for available slugs): tenant-<slug>-actions (open the kebab first), tenant-<slug>-suspend, tenant-<slug>-activate, tenant-<slug>-deactivate. Recipe to suspend a tenant: click nav-tenants, wait_for tenants-refresh, click tenant-<slug>-actions, wait_for tenant-<slug>-suspend, click tenant-<slug>-suspend.
 Login page: login-email, login-password, login-submit
 Notifications popover: appbar-notifications (open), notif-mark-all-read, notif-clear, notif-open-<id>, notif-dismiss-<id>
 Create a notification (hidden bridge — always available, even when not visible): fill the four hidden inputs, then click the submit button. Recipe:
@@ -77,6 +80,31 @@ wait_for assistant-ui-control
 \`\`\`"`
 
 
+import { formatAdminContextForPrompt } from "~/lib/admin-context"
+import { getOpenAITools, runLLMToolCalls } from "~/lib/admin-tools"
+import { ARCADIA_KNOWLEDGE } from "~/lib/arcadia-knowledge"
+import { useArcadiaClient } from "@crema/arcadia-client"
+
+/**
+ * Always includes domain knowledge + tools + admin context + persona.
+ * Adds the UI-control DSL/action catalog only when uiControl is on (those rules
+ * are about driving the cursor, irrelevant when the assistant is in plain Q&A).
+ */
+function buildAdminPreface(activeAgent: Agent | undefined, uiControl: boolean): string {
+  const persona = activeAgent
+    ? `Active persona: ${activeAgent.name} — ${activeAgent.role}\n${activeAgent.prompt}`
+    : ""
+  const ctx = formatAdminContextForPrompt()
+  const parts = [
+    "You are the operator's assistant inside Arcadia Admin. Be precise and direct. You have native function tools attached to this conversation — call them whenever the user asks about live platform state (counts, statuses, listings, lookups). Never invent tenant slugs, user counts, or statuses; if you need data, call a tool.",
+    ARCADIA_KNOWLEDGE,
+    persona,
+    ctx,
+  ]
+  if (uiControl) parts.push(UI_CONTROL_PREFACE)
+  return parts.filter(Boolean).join("\n\n")
+}
+
 function withTimeout<T>(p: Promise<T>, ms: number, signal: AbortSignal): Promise<T> {
   return new Promise<T>((resolve, reject) => {
     const t = setTimeout(() => {
@@ -395,8 +423,17 @@ function AssistantSurface({
     return () => clearTimeout(t)
   }, [uiControl])
 
-  const { messages, send, abort, isStreaming, error, reset } = useChat({
-    system: systemPrompt,
+  // Always-on admin system prompt — domain primer + tools + persona, regardless
+  // of UI Control. Per-call `send(text, {system})` overrides may not always be
+  // honored across hook re-renders, so anchor it at construction too.
+  const constructorSystemPrompt = buildSystemPrompt({
+    path: typeof window !== "undefined" ? window.location.pathname : "",
+    preface: buildAdminPreface(activeAgent, uiControl),
+    includeActions: uiControl,
+  })
+
+  const { messages, send, continueChat, abort, isStreaming, error, reset } = useChat({
+    system: constructorSystemPrompt,
     initialMessages: thread.messages as StoredMessage[],
   })
 
@@ -408,8 +445,21 @@ function AssistantSurface({
     const stamped: ThreadMessage[] = messages.map((m, i) => {
       const prior = thread.messages[i]
       if (m.role === "user") return { role: "user", content: m.content }
+      if (m.role === "tool") {
+        return {
+          role: "tool",
+          content: m.content,
+          toolCallId: m.toolCallId,
+          name: m.name,
+        }
+      }
       const agentId = prior?.agentId ?? activeAgentId
-      return { role: "assistant", content: m.content, agentId }
+      return {
+        role: "assistant",
+        content: m.content,
+        agentId,
+        ...(m.toolCalls ? { toolCalls: m.toolCalls } : {}),
+      }
     })
     updateThread(thread.id, {
       messages: stamped,
@@ -523,6 +573,50 @@ function AssistantSurface({
   }, [messages, onRequestRetry, thread.id, thread.pinned])
 
   const handleSendRef = useRef<((t: string) => void) | null>(null)
+  const arcadia = useArcadiaClient()
+  const toolIterationsRef = useRef(0)
+  const processedTurnRef = useRef(-1)
+  const prevStreamingRef = useRef(isStreaming)
+  const MAX_TOOL_ITERATIONS = 3
+
+  // Auto tool-loop using native function calls. When the streaming assistant
+  // turn ends with toolCalls attached, run them, push tool result messages,
+  // and continue the chat so the model produces its final answer.
+  useEffect(() => {
+    const justFinished = prevStreamingRef.current && !isStreaming
+    prevStreamingRef.current = isStreaming
+    if (!justFinished) return
+
+    const lastIdx = messages.length - 1
+    if (lastIdx < 0) return
+    const last = messages[lastIdx]
+    if (last.role !== "assistant") return
+    if (processedTurnRef.current === lastIdx) return
+    processedTurnRef.current = lastIdx
+
+    const calls = last.toolCalls ?? []
+    console.debug("[admin-tools] assistant turn finished, tool_calls:", calls)
+    if (calls.length === 0) {
+      toolIterationsRef.current = 0
+      return
+    }
+    if (toolIterationsRef.current >= MAX_TOOL_ITERATIONS) {
+      console.warn("[admin-tools] tool-iteration cap reached, dropping calls", calls)
+      return
+    }
+
+    toolIterationsRef.current += 1
+    void (async () => {
+      console.debug("[admin-tools] running tool calls", calls)
+      const { results, toolMessages } = await runLLMToolCalls(calls, { arcadia })
+      console.debug("[admin-tools] tool results", results)
+      void continueChat(toolMessages, {
+        system: constructorSystemPrompt,
+        tools: getOpenAITools(),
+        maxTokens: responseBudget,
+      })
+    })()
+  }, [messages, isStreaming, arcadia, continueChat, constructorSystemPrompt, responseBudget])
   const [editingIndex, setEditingIndex] = useState<number | null>(null)
   const [editingDraft, setEditingDraft] = useState("")
   const [speakingIndex, setSpeakingIndex] = useState<number | null>(null)
@@ -597,18 +691,12 @@ function AssistantSurface({
   }, [thread.id])
 
   const composedSystemPrompt = useMemo(() => {
-    return uiControl
-      ? buildSystemPrompt({
-          path:
-            typeof window !== "undefined" ? window.location.pathname : "",
-          preface: `${UI_CONTROL_PREFACE}\n\n${
-            activeAgent
-              ? `Active persona: ${activeAgent.name} — ${activeAgent.role}\n${activeAgent.prompt}`
-              : ""
-          }`,
-        })
-      : systemPrompt
-  }, [uiControl, systemPrompt, activeAgent])
+    return buildSystemPrompt({
+      path: typeof window !== "undefined" ? window.location.pathname : "",
+      preface: buildAdminPreface(activeAgent, uiControl),
+      includeActions: uiControl,
+    })
+  }, [uiControl, activeAgent])
 
   const buildTranscript = useCallback(() => {
     const lines: string[] = [
@@ -768,22 +856,14 @@ function AssistantSurface({
   }, [isStreaming, messages, uiControl])
 
   const handleSend: (text: string) => void = (text: string) => {
-    const system = uiControl
-      ? buildSystemPrompt({
-          path: window.location.pathname,
-          preface: `${UI_CONTROL_PREFACE}\n\n${activeAgent ? `Active persona: ${activeAgent.name} — ${activeAgent.role}\n${activeAgent.prompt}` : ""}`,
-        })
-      : systemPrompt
-    const sysTokens = estimateTokens(system)
-    const historyBudget = Math.max(
-      256,
-      contextTokens - sysTokens - responseBudget,
-    )
-    const userMsg = { role: "user" as const, content: text }
-    const trimmed = trimMessages([...messages, userMsg], historyBudget)
+    const system = buildSystemPrompt({
+      path: window.location.pathname,
+      preface: buildAdminPreface(activeAgent, uiControl),
+      includeActions: uiControl,
+    })
     void send(text, {
       system,
-      messages: trimmed,
+      tools: getOpenAITools(),
       maxTokens: responseBudget,
     })
   }
@@ -800,16 +880,15 @@ function AssistantSurface({
   }, [])
 
   const usedTokens = useMemo(() => {
-    const system = uiControl
-      ? buildSystemPrompt({
-          path: typeof window !== "undefined" ? window.location.pathname : "",
-          preface: UI_CONTROL_PREFACE,
-        })
-      : systemPrompt
+    const system = buildSystemPrompt({
+      path: typeof window !== "undefined" ? window.location.pathname : "",
+      preface: buildAdminPreface(activeAgent, uiControl),
+      includeActions: uiControl,
+    })
     const sysT = estimateTokens(system)
     const histT = messages.reduce((n, m) => n + estimateTokens(m.content), 0)
     return sysT + histT
-  }, [messages, uiControl, systemPrompt])
+  }, [messages, uiControl, systemPrompt, activeAgent])
 
   const suggestions = uiControl
     ? [
@@ -946,18 +1025,22 @@ function AssistantSurface({
                     ) : (
                       <div
                         className={
-                          isUser
-                            ? "rounded-2xl rounded-br-md bg-primary px-3.5 py-2 text-sm leading-relaxed text-primary-foreground"
-                            : "rounded-2xl rounded-bl-md bg-muted px-3.5 py-2 text-sm leading-relaxed text-foreground"
+                          m.role === "tool"
+                            ? "text-sm leading-relaxed"
+                            : isUser
+                              ? "rounded-2xl rounded-br-md bg-primary px-3.5 py-2 text-sm leading-relaxed text-primary-foreground"
+                              : "rounded-2xl rounded-bl-md bg-muted px-3.5 py-2 text-sm leading-relaxed text-foreground"
                         }
                       >
-                        {m.content ? (
+                        {m.role === "tool" ? (
+                          <MessageBody content={m.content} isToolResult />
+                        ) : m.content || (m.role === "assistant" && m.toolCalls?.length) ? (
                           isUser ? (
                             <span className="whitespace-pre-wrap">
                               {m.content}
                             </span>
                           ) : (
-                            <MessageBody content={m.content} />
+                            <MessageBody content={m.content} toolCalls={m.toolCalls} />
                           )
                         ) : isStreaming && i === messages.length - 1 ? (
                           "…"
diff --git a/app/routes/tenants.tsx b/app/routes/tenants.tsx
index 749f207..27e3413 100644
--- a/app/routes/tenants.tsx
+++ b/app/routes/tenants.tsx
@@ -36,6 +36,7 @@ import {
 } from "~/lib/arcadia/tenants"
 import { pageTitle } from "~/lib/page-meta"
 import { useSession } from "~/lib/session"
+import { useRegisterAdminContext } from "~/lib/admin-context"
 
 export const meta = () => pageTitle("Tenants")
 
@@ -131,13 +132,36 @@ export default function TenantsRoute() {
         header: "",
         align: "right",
         cell: (t) => (
-          <ActionsCell items={rowActions(t, arcadia, refresh, setPending, setError)} />
+          <ActionsCell
+            items={rowActions(t, arcadia, refresh, setPending, setError)}
+            triggerDataAction={`tenant-${t.slug}-actions`}
+          />
         ),
       },
     ],
     [arcadia, refresh],
   )
 
+  const tenantSummary = useMemo(
+    () => ({
+      total: tenants.length,
+      byStatus: tenants.reduce<Record<string, number>>((acc, t) => {
+        acc[t.status] = (acc[t.status] ?? 0) + 1
+        return acc
+      }, {}),
+      tenants: tenants.map((t) => ({
+        id: t.id,
+        slug: t.slug,
+        name: t.name,
+        status: t.status,
+        plan: t.plan?.name ?? null,
+        inserted_at: t.inserted_at,
+      })),
+    }),
+    [tenants],
+  )
+  useRegisterAdminContext("tenants", tenantSummary)
+
   const table = useTable<Tenant>({
     data: tenants,
     columns,
@@ -304,6 +328,7 @@ function rowActions(
       id: "suspend",
       label: "Suspend",
       icon: <Pause className="size-4" />,
+      dataAction: `tenant-${t.slug}-suspend`,
       onSelect: () => setPending({ kind: "suspend", tenant: t }),
     })
   } else {
@@ -311,6 +336,7 @@ function rowActions(
       id: "activate",
       label: "Activate",
       icon: <Play className="size-4" />,
+      dataAction: `tenant-${t.slug}-activate`,
       onSelect: async () => {
         try {
           await activateTenant(arcadia, t.id)
@@ -325,6 +351,7 @@ function rowActions(
     id: "deactivate",
     label: "Deactivate",
     destructive: true,
+    dataAction: `tenant-${t.slug}-deactivate`,
     onSelect: () => setPending({ kind: "deactivate", tenant: t }),
   })
   return items