diff --git a/internal/dockeragent/agent.go b/internal/dockeragent/agent.go
index c448ddcc6..8fa3435eb 100644
--- a/internal/dockeragent/agent.go
+++ b/internal/dockeragent/agent.go
@@ -76,28 +76,27 @@ const (
 
 // Agent collects Docker metrics and posts them to Pulse.
 type Agent struct {
-	cfg                 Config
-	docker              dockerClient
-	daemonHost          string
-	daemonID            string // Cached at init; Podman can return unstable IDs across calls
-	runtime             RuntimeKind
-	runtimeVer          string
-	agentVersion        string
-	supportsSwarm       bool
-	httpClients         map[bool]*http.Client
-	logger              zerolog.Logger
-	machineID           string
-	hostName            string
-	cpuCount            int
-	targets             []TargetConfig
-	allowedStates       map[string]struct{}
-	stateFilters        []string
-	hostID              string
-	prevContainerCPU    map[string]cpuSample
-	cpuMu               sync.Mutex // protects prevContainerCPU and preCPUStatsFailures
-	preCPUStatsFailures int
-	reportBuffer        *buffer.Queue[agentsdocker.Report]
-	registryChecker     *RegistryChecker // For checking container image updates
+	cfg              Config
+	docker           dockerClient
+	daemonHost       string
+	daemonID         string // Cached at init; Podman can return unstable IDs across calls
+	runtime          RuntimeKind
+	runtimeVer       string
+	agentVersion     string
+	supportsSwarm    bool
+	httpClients      map[bool]*http.Client
+	logger           zerolog.Logger
+	machineID        string
+	hostName         string
+	cpuCount         int
+	targets          []TargetConfig
+	allowedStates    map[string]struct{}
+	stateFilters     []string
+	hostID           string
+	prevContainerCPU map[string]cpuSample
+	cpuMu            sync.Mutex // protects prevContainerCPU
+	reportBuffer     *buffer.Queue[agentsdocker.Report]
+	registryChecker  *RegistryChecker // For checking container image updates
 }
 
 // ErrStopRequested indicates the agent should terminate gracefully after acknowledging a stop command.
diff --git a/internal/dockeragent/agent_cpu_test.go b/internal/dockeragent/agent_cpu_test.go
index 2c88b3eb6..9224e8b35 100644
--- a/internal/dockeragent/agent_cpu_test.go
+++ b/internal/dockeragent/agent_cpu_test.go
@@ -11,32 +11,42 @@ import (
 func TestCalculateContainerCPUPercent(t *testing.T) {
 	logger := zerolog.Nop()
 
-	t.Run("uses precpu stats", func(t *testing.T) {
+	t.Run("manual delta across two calls", func(t *testing.T) {
 		agent := &Agent{
 			logger:           logger,
 			prevContainerCPU: make(map[string]cpuSample),
 			cpuCount:         2,
 		}
 
-		stats := containertypes.StatsResponse{
+		// First call: stores sample, returns 0
+		stats1 := containertypes.StatsResponse{
+			Read: time.Now().Add(-time.Second),
+			CPUStats: containertypes.CPUStats{
+				CPUUsage:    containertypes.CPUUsage{TotalUsage: 100},
+				SystemUsage: 1000,
+				OnlineCPUs:  2,
+			},
+		}
+		got := agent.calculateContainerCPUPercent("container-123456", stats1)
+		if got != 0 {
+			t.Fatalf("first call: expected 0, got %f", got)
+		}
+		if _, ok := agent.prevContainerCPU["container-123456"]; !ok {
+			t.Fatal("expected sample to be stored after first call")
+		}
+
+		// Second call: computes delta from stored sample
+		stats2 := containertypes.StatsResponse{
 			Read: time.Now(),
 			CPUStats: containertypes.CPUStats{
 				CPUUsage:    containertypes.CPUUsage{TotalUsage: 200},
 				SystemUsage: 2000,
 				OnlineCPUs:  2,
 			},
-			PreCPUStats: containertypes.CPUStats{
-				CPUUsage:    containertypes.CPUUsage{TotalUsage: 100},
-				SystemUsage: 1000,
-			},
 		}
-
-		got := agent.calculateContainerCPUPercent("container-123456", stats)
+		got = agent.calculateContainerCPUPercent("container-123456", stats2)
 		if got <= 0 {
-			t.Fatalf("expected percent > 0, got %f", got)
-		}
-		if _, ok := agent.prevContainerCPU["container-123456"]; !ok {
-			t.Fatal("expected current sample to be stored")
+			t.Fatalf("second call: expected percent > 0, got %f", got)
 		}
 	})
 
@@ -286,17 +296,21 @@ func TestCalculateContainerCPUPercent(t *testing.T) {
 		}
 	})
 
-	t.Run("warn after repeated failures", func(t *testing.T) {
+	t.Run("ignores stale precpu stats", func(t *testing.T) {
 		agent := &Agent{
 			logger:           logger,
 			prevContainerCPU: make(map[string]cpuSample),
+			cpuCount:         2,
 		}
 
+		// Even with valid-looking PreCPUStats, we use manual tracking.
+		// First call stores sample and returns 0.
 		stats := containertypes.StatsResponse{
 			Read: time.Now(),
 			CPUStats: containertypes.CPUStats{
-				CPUUsage:    containertypes.CPUUsage{TotalUsage: 100},
-				SystemUsage: 1000,
+				CPUUsage:    containertypes.CPUUsage{TotalUsage: 200},
+				SystemUsage: 2000,
+				OnlineCPUs:  2,
 			},
 			PreCPUStats: containertypes.CPUStats{
 				CPUUsage:    containertypes.CPUUsage{TotalUsage: 100},
@@ -304,8 +318,9 @@ func TestCalculateContainerCPUPercent(t *testing.T) {
 			},
 		}
 
-		for i := 0; i < 10; i++ {
-			_ = agent.calculateContainerCPUPercent("container-123456", stats)
+		got := agent.calculateContainerCPUPercent("container-123456", stats)
+		if got != 0 {
+			t.Fatalf("expected 0 on first call (manual tracking), got %f", got)
 		}
 	})
 }
diff --git a/internal/dockeragent/collect.go b/internal/dockeragent/collect.go
index a2579b953..664e4a4ae 100644
--- a/internal/dockeragent/collect.go
+++ b/internal/dockeragent/collect.go
@@ -605,24 +605,11 @@ func (a *Agent) calculateContainerCPUPercent(id string, stats containertypes.Sta
 		read:        stats.Read,
 	}
 
-	// Try to use PreCPUStats if available
-	percent := calculateCPUPercent(stats, a.cpuCount)
-	if percent > 0 {
-		a.prevContainerCPU[id] = current
-		a.logger.Debug().
-			Str("container_id", id[:12]).
-			Float64("cpu_percent", percent).
-			Msg("CPU calculated from PreCPUStats")
-		return percent
-	}
-
-	// PreCPUStats not available or invalid, use manual tracking
-	a.preCPUStatsFailures++
-	if a.preCPUStatsFailures == 10 {
-		a.logger.Warn().
-			Str("runtime", string(a.runtime)).
-			Msg("PreCPUStats consistently unavailable from Docker API - using manual CPU tracking (this is normal for one-shot stats)")
-	}
+	// Always use manual delta tracking. Docker's PreCPUStats is unreliable for
+	// one-shot stats because many Docker versions don't update their internal
+	// cache between non-streaming reads, causing PreCPUStats to remain stale
+	// (from container start) and producing a constant lifetime-average CPU%
+	// instead of a current value.
 	prev, ok := a.prevContainerCPU[id]
 	if !ok {
 		// First time seeing this container - store current sample and return 0