diff --git a/cmd/ai.go b/cmd/ai.go
index 0e74be10..e6bb47fa 100644
--- a/cmd/ai.go
+++ b/cmd/ai.go
@@ -499,9 +499,6 @@ func handleAICopilot(r *fastglue.Request) error {
 	history = append(history, aimodels.ChatMessage{Role: aimodels.RoleUser, Content: req.Message})
 
 	scope := ai.AgentRunScope{AgentID: auser.ID, ConversationID: conv.ID, ConversationUUID: conv.UUID, Surface: aimodels.ToolInvocationCopilot, UserMessage: req.Message}
-	if _, approval := app.ai.PendingAgentApproval(scope); approval != nil {
-		return sendErrorEnvelope(r, envelope.NewError(envelope.ConflictError, app.i18n.T("ai.toolApprovalPending"), nil))
-	}
 	toolIDs, err := app.ai.GetEnabledCopilotToolIDs()
 	if err != nil {
 		return sendErrorEnvelope(r, err)
diff --git a/i18n/en-US.json b/i18n/en-US.json
index 6de71e61..9eda9e75 100644
--- a/i18n/en-US.json
+++ b/i18n/en-US.json
@@ -602,7 +602,7 @@
   "ai.noTagsConfigured": "No tags exist yet. Tag suggestions need an existing tag list to choose from. Ask an admin to create tags first.",
   "ai.tagSuggestEmptyConversation": "There are no messages to suggest tags from yet.",
   "ai.toolApprovalExpired": "This tool request is no longer available. Run the AI action again to create a new request.",
-  "ai.toolApprovalPending": "A tool request is waiting for approval. Approve or decline it before sending another Copilot message.",
+  "ai.toolApprovalPending": "A tool request is waiting for approval. The AI cannot run again until it is reviewed. Approve or decline the pending request first.",
   "ai.toolApprovalUnavailable": "This tool request cannot be approved for this conversation. Run the AI action again to create a new request.",
   "ai.toolChangedBeforeApproval": "This tool changed after approval was requested and was not run. Run the AI action again to review the current tool.",
   "ai.toolApprovalTitle": "Review tool request",
diff --git a/internal/ai/agent_approval.go b/internal/ai/agent_approval.go
index 9adae8b0..3e5dcb06 100644
--- a/internal/ai/agent_approval.go
+++ b/internal/ai/agent_approval.go
@@ -118,7 +118,11 @@ func (m *Manager) ClearPendingAgentRuns(scope AgentRunScope) {
 }
 
 func (m *Manager) continueAgentRun(ctx context.Context, run *pendingAgentRun) (models.AgentRunResult, error) {
-	if result := m.processAgentToolCalls(ctx, run); result != nil {
+	result, err := m.processAgentToolCalls(ctx, run)
+	if err != nil {
+		return models.AgentRunResult{}, err
+	}
+	if result != nil {
 		return *result, nil
 	}
 	for run.Step < run.MaxSteps {
@@ -136,7 +140,11 @@ func (m *Manager) continueAgentRun(ctx context.Context, run *pendingAgentRun) (m
 		run.ToolCalls = res.ToolCalls
 		run.NextToolCall = 0
 		run.Step++
-		if result := m.processAgentToolCalls(ctx, run); result != nil {
+		result, err := m.processAgentToolCalls(ctx, run)
+		if err != nil {
+			return models.AgentRunResult{}, err
+		}
+		if result != nil {
 			return *result, nil
 		}
 	}
@@ -153,7 +161,7 @@ func (m *Manager) continueAgentRun(ctx context.Context, run *pendingAgentRun) (m
 	return completedAgentRun(res.Content), nil
 }
 
-func (m *Manager) processAgentToolCalls(ctx context.Context, run *pendingAgentRun) *models.AgentRunResult {
+func (m *Manager) processAgentToolCalls(ctx context.Context, run *pendingAgentRun) (*models.AgentRunResult, error) {
 	for run.NextToolCall < len(run.ToolCalls) {
 		toolCall := run.ToolCalls[run.NextToolCall]
 		if run.Approvals {
@@ -163,15 +171,17 @@ func (m *Manager) processAgentToolCalls(ctx context.Context, run *pendingAgentRu
 				run.ToolUpdatedAt = customTool.toolUpdatedAt()
 				run.CreatedAt = time.Now()
 				run.ExpiresAt = run.CreatedAt.Add(pendingAgentRunTTL)
-				m.storeAgentRun(run)
-				return &models.AgentRunResult{Status: models.AgentRunApprovalRequired, Approval: agentRunApproval(run)}
+				if err := m.storeAgentRun(run); err != nil {
+					return nil, err
+				}
+				return &models.AgentRunResult{Status: models.AgentRunApprovalRequired, Approval: agentRunApproval(run)}, nil
 			}
 		}
 		m.appendToolResult(run, toolCall, m.executeToolCall(ctx, run.Registry, toolCall))
 		run.NextToolCall++
 	}
 	run.ToolCalls = nil
-	return nil
+	return nil, nil
 }
 
 func (m *Manager) appendToolResult(run *pendingAgentRun, toolCall models.ToolCall, result string) {
@@ -183,11 +193,19 @@ func (m *Manager) appendToolResult(run *pendingAgentRun, toolCall models.ToolCal
 	})
 }
 
-func (m *Manager) storeAgentRun(run *pendingAgentRun) {
+// storeAgentRun admits one pending run per scope, so two requests racing on the same
+// conversation and surface cannot both park a tool call the agent could approve twice.
+func (m *Manager) storeAgentRun(run *pendingAgentRun) error {
 	m.pendingRunMu.Lock()
 	defer m.pendingRunMu.Unlock()
 	m.deleteExpiredAgentRunsLocked(time.Now())
+	for _, pending := range m.pendingRuns {
+		if sameAgentRunScope(pending.Scope, run.Scope) {
+			return envelope.NewError(envelope.ConflictError, m.i18n.T("ai.toolApprovalPending"), nil)
+		}
+	}
 	m.pendingRuns[run.ID] = run
+	return nil
 }
 
 func (m *Manager) claimAgentRun(runID string, agentID int) (*pendingAgentRun, error) {
diff --git a/internal/ai/agent_approval_test.go b/internal/ai/agent_approval_test.go
index f67d5884..450bb1cc 100644
--- a/internal/ai/agent_approval_test.go
+++ b/internal/ai/agent_approval_test.go
@@ -124,6 +124,33 @@ func TestAgentToolDecisionIsFinalWhenFollowUpFails(t *testing.T) {
 	}
 }
 
+func TestSecondRunInSameScopeIsRejectedWhileOneAwaitsApproval(t *testing.T) {
+	m := newTestManager(t)
+	tool := &approvalTestTool{}
+	scope := AgentRunScope{AgentID: 7, ConversationID: 9, ConversationUUID: "conversation-1", Surface: models.ToolInvocationCopilot}
+	newRun := func() *pendingAgentRun {
+		return &pendingAgentRun{
+			Scope: scope,
+			Client: &approvalTestProvider{responses: []models.ChatCompletionResult{
+				{ToolCalls: []models.ToolCall{{ID: "call-1", Type: "function", Function: models.ToolCallFunction{Name: tool.Name(), Arguments: `{}`}}}},
+			}},
+			Registry:  map[string]Tool{tool.Name(): tool},
+			MaxSteps:  5,
+			Approvals: true,
+		}
+	}
+
+	if _, err := m.continueAgentRun(t.Context(), newRun()); err != nil {
+		t.Fatalf("first run: %v", err)
+	}
+	if _, err := m.continueAgentRun(t.Context(), newRun()); err == nil {
+		t.Fatal("expected the second run in the same scope to be rejected")
+	}
+	if len(m.pendingRuns) != 1 {
+		t.Fatalf("expected one pending run, got %d", len(m.pendingRuns))
+	}
+}
+
 func (t *approvalTestTool) approvalRequired() bool { return true }
 
 func (t *approvalTestTool) toolID() int { return 1 }