diff --git a/README.md b/README.md
index 375d8a7..2630bce 100644
--- a/README.md
+++ b/README.md
@@ -203,6 +203,7 @@ agk trace mermaid > trace_flow.md
 | `trace list` | List all captured trace runs. |
 | `trace show` | Display summary of a specific run. |
 | `trace view` | Open the interactive TUI trace explorer. |
+| `trace diff` | Compare two trace runs (duration, tokens, cost, LLM calls). |
 | `trace mermaid` | Generate Mermaid flowchart of trace execution. |
 
 ---
diff --git a/cmd/trace_diff.go b/cmd/trace_diff.go
new file mode 100644
index 0000000..4fc72d2
--- /dev/null
+++ b/cmd/trace_diff.go
@@ -0,0 +1,202 @@
+package cmd
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"text/tabwriter"
+	"time"
+
+	"github.com/fatih/color"
+	"github.com/spf13/cobra"
+)
+
+// diffMetric is a single comparable metric between two runs.
+type diffMetric struct {
+	Label       string
+	A, B        float64
+	LowerBetter bool             // a lower B than A is an improvement
+	Colorize    bool             // whether the delta should be colored good/bad
+	Format      func(float64) string
+}
+
+// diffCmd compares two trace runs to answer "did my change help?".
+var diffCmd = &cobra.Command{
+	Use:   "diff [run-a] [run-b]",
+	Short: "Compare two trace runs (duration, tokens, cost, ...)",
+	Long: `Compare two trace runs and show the deltas for duration, spans, LLM calls,
+tokens, and estimated cost.
+
+Run selection:
+  agk trace diff                 # compare the two most recent runs
+  agk trace diff <run-a>         # compare <run-a> (baseline) against the latest run
+  agk trace diff <run-a> <run-b> # compare two explicit runs
+
+The first run is treated as the baseline (A); the second is the new run (B).
+For duration, tokens, and cost, a lower value in B is shown as an improvement.`,
+	Args: cobra.MaximumNArgs(2),
+	RunE: func(cmd *cobra.Command, args []string) error {
+		return runTraceDiff(args)
+	},
+}
+
+func init() {
+	traceCmd.AddCommand(diffCmd)
+}
+
+func runTraceDiff(args []string) error {
+	idA, idB, err := resolveDiffRuns(args)
+	if err != nil {
+		return err
+	}
+
+	for _, id := range []string{idA, idB} {
+		if _, err := os.Stat(filepath.Join(runsDirName, id)); err != nil {
+			return fmt.Errorf("trace not found: %s", id)
+		}
+	}
+
+	manifestA, err := readManifest(filepath.Join(runsDirName, idA))
+	if err != nil {
+		return fmt.Errorf("failed to read run %s: %w", idA, err)
+	}
+	manifestB, err := readManifest(filepath.Join(runsDirName, idB))
+	if err != nil {
+		return fmt.Errorf("failed to read run %s: %w", idB, err)
+	}
+
+	printDiff(manifestA, manifestB)
+	return nil
+}
+
+// resolveDiffRuns determines the two run IDs to compare based on the args provided.
+func resolveDiffRuns(args []string) (string, string, error) {
+	switch len(args) {
+	case 2:
+		return args[0], args[1], nil
+	case 1:
+		latest := ""
+		for _, id := range recentRunIDs(runsDirName) {
+			if id != args[0] {
+				latest = id
+				break
+			}
+		}
+		if latest == "" {
+			return "", "", fmt.Errorf("need a second run to diff against %s", args[0])
+		}
+		return args[0], latest, nil
+	default: // 0 args
+		ids := recentRunIDs(runsDirName)
+		if len(ids) < 2 {
+			return "", "", fmt.Errorf("need at least two trace runs to diff (found %d)", len(ids))
+		}
+		// ids[0] is newest; baseline is the older of the two most recent.
+		return ids[1], ids[0], nil
+	}
+}
+
+// recentRunIDs returns run directory names sorted newest-first by modification time.
+func recentRunIDs(runsDir string) []string {
+	entries, err := os.ReadDir(runsDir)
+	if err != nil {
+		return nil
+	}
+	type run struct {
+		name string
+		mod  time.Time
+	}
+	var runs []run
+	for _, e := range entries {
+		if !e.IsDir() {
+			continue
+		}
+		info, err := e.Info()
+		if err != nil {
+			continue
+		}
+		runs = append(runs, run{e.Name(), info.ModTime()})
+	}
+	sort.Slice(runs, func(i, j int) bool { return runs[i].mod.After(runs[j].mod) })
+
+	ids := make([]string, len(runs))
+	for i, r := range runs {
+		ids[i] = r.name
+	}
+	return ids
+}
+
+func runDiffMetrics(a, b TraceRun) []diffMetric {
+	return []diffMetric{
+		{Label: "Duration", A: a.Duration, B: b.Duration, LowerBetter: true, Colorize: true, Format: fmtSeconds},
+		{Label: "Spans", A: float64(a.SpanCount), B: float64(b.SpanCount), Format: fmtCount},
+		{Label: "LLM Calls", A: float64(a.LLMCalls), B: float64(b.LLMCalls), Format: fmtCount},
+		{Label: "Tokens", A: float64(a.TotalTokens), B: float64(b.TotalTokens), LowerBetter: true, Colorize: true, Format: fmtCount},
+		{Label: "Est. Cost", A: a.EstimatedCost, B: b.EstimatedCost, LowerBetter: true, Colorize: true, Format: fmtUSD},
+	}
+}
+
+// deltaDirection returns +1 if B is an improvement over A, -1 if a regression,
+// and 0 if unchanged or not colorized.
+func deltaDirection(m diffMetric) int {
+	d := m.B - m.A
+	if d == 0 || !m.Colorize {
+		return 0
+	}
+	if (d < 0) == m.LowerBetter {
+		return 1
+	}
+	return -1
+}
+
+func printDiff(a, b TraceRun) {
+	fmt.Println()
+	color.Cyan("📊 Trace Diff")
+	fmt.Printf("  A (baseline): %s\n", a.RunID)
+	fmt.Printf("  B (new):      %s\n", b.RunID)
+	fmt.Println(strings.Repeat("─", 64))
+
+	w := tabwriter.NewWriter(os.Stdout, 0, 0, 2, ' ', 0)
+	fmt.Fprintln(w, "METRIC\tA (baseline)\tB (new)\tΔ")
+	for _, m := range runDiffMetrics(a, b) {
+		delta := formatDelta(m)
+		switch deltaDirection(m) {
+		case 1:
+			delta = color.GreenString(delta)
+		case -1:
+			delta = color.RedString(delta)
+		}
+		fmt.Fprintf(w, "%s\t%s\t%s\t%s\n", m.Label, m.Format(m.A), m.Format(m.B), delta)
+	}
+	_ = w.Flush()
+	fmt.Println()
+}
+
+func formatDelta(m diffMetric) string {
+	d := m.B - m.A
+	if d == 0 {
+		return "—"
+	}
+	sign, arrow := "+", "▲"
+	if d < 0 {
+		sign, arrow = "-", "▼"
+	}
+	out := fmt.Sprintf("%s%s %s", sign, m.Format(absF(d)), arrow)
+	if m.A != 0 {
+		out += fmt.Sprintf(" %.0f%%", d/m.A*100)
+	}
+	return out
+}
+
+func fmtSeconds(v float64) string { return fmt.Sprintf("%.2fs", v) }
+func fmtCount(v float64) string   { return fmt.Sprintf("%.0f", v) }
+func fmtUSD(v float64) string     { return fmt.Sprintf("$%.4f", v) }
+
+func absF(v float64) float64 {
+	if v < 0 {
+		return -v
+	}
+	return v
+}
diff --git a/cmd/trace_diff_test.go b/cmd/trace_diff_test.go
new file mode 100644
index 0000000..117e09b
--- /dev/null
+++ b/cmd/trace_diff_test.go
@@ -0,0 +1,73 @@
+package cmd
+
+import "testing"
+
+func TestDeltaDirection(t *testing.T) {
+	cases := []struct {
+		name string
+		m    diffMetric
+		want int
+	}{
+		{"lower-better improvement", diffMetric{A: 10, B: 5, LowerBetter: true, Colorize: true}, 1},
+		{"lower-better regression", diffMetric{A: 5, B: 10, LowerBetter: true, Colorize: true}, -1},
+		{"equal is neutral", diffMetric{A: 7, B: 7, LowerBetter: true, Colorize: true}, 0},
+		{"not colorized is neutral", diffMetric{A: 10, B: 5, LowerBetter: true, Colorize: false}, 0},
+		{"higher-better improvement", diffMetric{A: 5, B: 10, LowerBetter: false, Colorize: true}, 1},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			if got := deltaDirection(c.m); got != c.want {
+				t.Errorf("deltaDirection = %d, want %d", got, c.want)
+			}
+		})
+	}
+}
+
+func TestRunDiffMetrics(t *testing.T) {
+	a := TraceRun{Duration: 2.0, SpanCount: 5, LLMCalls: 2, TotalTokens: 1000, EstimatedCost: 0.0100}
+	b := TraceRun{Duration: 1.0, SpanCount: 4, LLMCalls: 1, TotalTokens: 500, EstimatedCost: 0.0050}
+
+	metrics := runDiffMetrics(a, b)
+	if len(metrics) != 5 {
+		t.Fatalf("expected 5 metrics, got %d", len(metrics))
+	}
+
+	byLabel := make(map[string]diffMetric, len(metrics))
+	for _, m := range metrics {
+		byLabel[m.Label] = m
+	}
+
+	// Tokens halved → improvement.
+	if dir := deltaDirection(byLabel["Tokens"]); dir != 1 {
+		t.Errorf("Tokens direction = %d, want 1 (improvement)", dir)
+	}
+	// Cost halved → improvement.
+	if dir := deltaDirection(byLabel["Est. Cost"]); dir != 1 {
+		t.Errorf("Cost direction = %d, want 1 (improvement)", dir)
+	}
+	// Spans is not colorized → neutral even though it changed.
+	if dir := deltaDirection(byLabel["Spans"]); dir != 0 {
+		t.Errorf("Spans direction = %d, want 0 (neutral)", dir)
+	}
+}
+
+func TestFormatDelta(t *testing.T) {
+	if got := formatDelta(diffMetric{A: 10, B: 10, Format: fmtCount}); got != "—" {
+		t.Errorf("equal delta = %q, want em dash", got)
+	}
+	// 1000 -> 500 tokens: -500, 50% lower.
+	got := formatDelta(diffMetric{A: 1000, B: 500, Format: fmtCount})
+	if got != "-500 ▼ -50%" {
+		t.Errorf("formatDelta = %q, want %q", got, "-500 ▼ -50%")
+	}
+}
+
+func TestResolveDiffRunsExplicit(t *testing.T) {
+	a, b, err := resolveDiffRuns([]string{"run-1", "run-2"})
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if a != "run-1" || b != "run-2" {
+		t.Errorf("got (%s, %s), want (run-1, run-2)", a, b)
+	}
+}