diff --git a/pkg/agent/loop_execute.go b/pkg/agent/loop_execute.go
index 1b2ee47..61f054f 100644
--- a/pkg/agent/loop_execute.go
+++ b/pkg/agent/loop_execute.go
@@ -26,6 +26,24 @@ func newToolCallID() string {
 // short-circuit with a synthesized tool-error result. A fatal error
 // recorded by any goroutine (e.g. anti-loop detector) breaks out of
 // the wave loop.
+// defaultMaxParallelToolCalls bounds concurrent tool execution within one
+// dependency wave. Wide fan-outs are the pathological case this exists for:
+// a model asking for 40 calls at once would otherwise open 40 sockets or
+// spawn 40 subprocesses. Waves are usually smaller than this, so the cap is
+// invisible in normal operation.
+const defaultMaxParallelToolCalls = 8
+
+// toolCallSemaphore returns a buffered channel bounding wave concurrency, or
+// nil when no cap applies — an unlimited setting, or a wave already at or
+// below the cap, both skip the channel and its send/receive pair entirely.
+func (al *AgentLoop) toolCallSemaphore(waveSize int) chan struct{} {
+	n := al.MaxParallelToolCalls
+	if n <= 0 || waveSize <= n {
+		return nil
+	}
+	return make(chan struct{}, n)
+}
+
 func (al *AgentLoop) executeToolWaves(ctx context.Context, st *iterationState, scheduled []PendingToolCall) *waveState {
 	waves, schedErr := scheduleToolCalls(scheduled)
 	if schedErr != nil {
@@ -43,11 +61,16 @@ func (al *AgentLoop) executeToolWaves(ctx context.Context, st *iterationState, s
 			}))
 		})
 
+		sem := al.toolCallSemaphore(len(substitutedWave))
 		var wg sync.WaitGroup
 		for _, tc := range substitutedWave {
 			wg.Add(1)
 			go func(tCall PendingToolCall) {
 				defer wg.Done()
+				if sem != nil {
+					sem <- struct{}{}
+					defer func() { <-sem }()
+				}
 				al.executeToolCall(ctx, st, ws, tCall)
 			}(tc)
 		}
diff --git a/pkg/agent/loop_parallel_cap_test.go b/pkg/agent/loop_parallel_cap_test.go
new file mode 100644
index 0000000..a86c83c
--- /dev/null
+++ b/pkg/agent/loop_parallel_cap_test.go
@@ -0,0 +1,123 @@
+package agent
+
+import (
+	"context"
+	"fmt"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/hung12ct/gopheragent/pkg/tools"
+)
+
+// concurrencyProbe records the high-water mark of overlapping Execute calls
+// so a test can assert the wave semaphore actually bounds them.
+type concurrencyProbe struct {
+	live atomic.Int32
+	peak atomic.Int32
+	runs atomic.Int32
+}
+
+func (c *concurrencyProbe) Descriptor() tools.ToolDescriptor {
+	return tools.ToolDescriptor{
+		Name:        "probe",
+		Description: "records concurrency",
+		Display:     tools.DefaultDisplay("probe", "records concurrency"),
+	}
+}
+
+func (c *concurrencyProbe) Execute(_ context.Context, args string) (tools.Result, error) {
+	live := c.live.Add(1)
+	for {
+		peak := c.peak.Load()
+		if live <= peak || c.peak.CompareAndSwap(peak, live) {
+			break
+		}
+	}
+	// Hold the slot long enough that an unbounded fan-out genuinely overlaps.
+	time.Sleep(5 * time.Millisecond)
+	c.live.Add(-1)
+	c.runs.Add(1)
+	return tools.Text("probe:" + args), nil
+}
+
+// distinctCalls builds n calls to "probe" with unique arguments, so neither
+// the anti-loop detector nor the result cache collapses them.
+func distinctCalls(n int) []PendingToolCall {
+	out := make([]PendingToolCall, n)
+	for i := range n {
+		out[i] = PendingToolCall{
+			ID:       fmt.Sprintf("p%d", i),
+			Name:     "probe",
+			ArgsJSON: fmt.Sprintf(`{"i":%d}`, i),
+		}
+	}
+	return out
+}
+
+func runProbeWave(t *testing.T, calls int, opts ...Option) *concurrencyProbe {
+	t.Helper()
+	probe := &concurrencyProbe{}
+	provider := &scriptProvider{turns: []LLMResult{
+		{ToolCalls: distinctCalls(calls)},
+		{Content: "final"},
+	}}
+	loop, _ := setup(provider, probe)
+	for _, opt := range opts {
+		opt(loop)
+	}
+
+	if _, err := loop.RunIteration(context.Background(), "s1", "go"); err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if got := probe.runs.Load(); got != int32(calls) {
+		t.Fatalf("executed %d calls, want %d — the cap must delay calls, never drop them", got, calls)
+	}
+	return probe
+}
+
+func TestMaxParallelToolCalls_BoundsWaveConcurrency(t *testing.T) {
+	probe := runProbeWave(t, 20, WithMaxParallelToolCalls(4))
+
+	if peak := probe.peak.Load(); peak > 4 {
+		t.Fatalf("peak concurrency = %d, want <= 4", peak)
+	}
+}
+
+func TestMaxParallelToolCalls_DefaultApplies(t *testing.T) {
+	probe := runProbeWave(t, 20)
+
+	if peak := probe.peak.Load(); peak > defaultMaxParallelToolCalls {
+		t.Fatalf("peak concurrency = %d, want <= %d (default)", peak, defaultMaxParallelToolCalls)
+	}
+}
+
+func TestMaxParallelToolCalls_ZeroIsUnlimited(t *testing.T) {
+	probe := runProbeWave(t, 20, WithMaxParallelToolCalls(0))
+
+	// Not asserting an exact peak: the scheduler is free to interleave. The
+	// point is that no cap is applied, which only shows up as a peak above
+	// the default the constructor would otherwise have installed.
+	if peak := probe.peak.Load(); peak <= defaultMaxParallelToolCalls {
+		t.Skipf("peak %d did not exceed the default cap; scheduling-dependent, not a failure", peak)
+	}
+}
+
+// toolCallSemaphore must skip the channel entirely when it cannot bind, so
+// the common small-wave path pays nothing.
+func TestToolCallSemaphore_NilWhenNoCapBinds(t *testing.T) {
+	al := &AgentLoop{MaxParallelToolCalls: 8}
+	for _, size := range []int{0, 1, 8} {
+		if sem := al.toolCallSemaphore(size); sem != nil {
+			t.Fatalf("waveSize %d: got a semaphore, want nil", size)
+		}
+	}
+	if sem := al.toolCallSemaphore(9); sem == nil || cap(sem) != 8 {
+		t.Fatalf("waveSize 9: want a semaphore of cap 8, got %v", sem)
+	}
+
+	unlimited := &AgentLoop{MaxParallelToolCalls: 0}
+	if sem := unlimited.toolCallSemaphore(100); sem != nil {
+		t.Fatal("unlimited: got a semaphore, want nil")
+	}
+}
diff --git a/pkg/agent/loop_stream.go b/pkg/agent/loop_stream.go
index d033bef..ac09b6e 100644
--- a/pkg/agent/loop_stream.go
+++ b/pkg/agent/loop_stream.go
@@ -426,6 +426,14 @@ type AgentLoop struct {
 	// they did not execute. A "thought" event announces the truncation.
 	MaxToolCallsPerTurn int
 
+	// MaxParallelToolCalls caps how many tool calls execute concurrently
+	// within one dependency wave. Defaults to defaultMaxParallelToolCalls;
+	// 0 means unlimited. Unlike MaxToolCallsPerTurn this drops nothing — a
+	// call over the cap waits for a slot, so the wave's result set is
+	// identical either way. It bounds live resource use (sockets,
+	// subprocesses, provider rate limits) when a model emits a wide fan-out.
+	MaxParallelToolCalls int
+
 	// MaxToolCallsPerSession caps the cumulative number of tool calls
 	// scheduled across all iterations of a single Run. 0 (default) means
 	// unlimited. Distinct from MaxToolCallsPerTurn, which only bounds the
@@ -518,13 +526,14 @@ type AgentLoop struct {
 // Optional hooks run before each iteration for security/policy enforcement.
 func NewAgentLoop(sessions SessionManager, registry *tools.Registry, llm LLMProvider, hooks ...Hook) *AgentLoop {
 	return &AgentLoop{
-		Sessions:        sessions,
-		Tools:           registry,
-		LLM:             llm,
-		MaxIters:        15,
-		EmitThoughts:    true,
-		BeforeHooks:     hooks,
-		AutoCacheSystem: true,
+		Sessions:             sessions,
+		Tools:                registry,
+		LLM:                  llm,
+		MaxIters:             15,
+		EmitThoughts:         true,
+		BeforeHooks:          hooks,
+		AutoCacheSystem:      true,
+		MaxParallelToolCalls: defaultMaxParallelToolCalls,
 	}
 }
 
diff --git a/pkg/agent/options.go b/pkg/agent/options.go
index 78b629c..6affbb3 100644
--- a/pkg/agent/options.go
+++ b/pkg/agent/options.go
@@ -245,6 +245,16 @@ func WithMaxToolCallsPerTurn(n int) Option {
 	return func(al *AgentLoop) { al.MaxToolCallsPerTurn = n }
 }
 
+// WithMaxParallelToolCalls caps how many tool calls run concurrently inside
+// one dependency wave. Defaults to 8; pass 0 for unlimited. Nothing is
+// dropped — calls over the cap wait for a slot, so results are unchanged and
+// only peak resource use differs. Raise it for latency-bound tools (HTTP
+// fan-out), lower it for tools that each hold a scarce resource (database
+// connections, subprocesses).
+func WithMaxParallelToolCalls(n int) Option {
+	return func(al *AgentLoop) { al.MaxParallelToolCalls = n }
+}
+
 // WithMaxToolCallsPerSession caps cumulative tool calls across all
 // iterations of a single Run. 0 (default) means unlimited. When the cap
 // trips, the loop emits LimitExhaustedEvent and saves history.