diff --git a/pkg/agentic/queue.go b/pkg/agentic/queue.go
index 52a5669..9aa04d1 100644
--- a/pkg/agentic/queue.go
+++ b/pkg/agentic/queue.go
@@ -32,14 +32,49 @@ type RateConfig struct {
 	BurstDelay     int    `yaml:"burst_delay"`     // Delay during burst window
 }
 
+// ConcurrencyLimit supports both flat (int) and nested (map with total + per-model) formats.
+//
+//	claude: 1                       → Total=1, Models=nil
+//	codex:                          → Total=2, Models={"gpt-5.4": 1, "gpt-5.3-codex-spark": 1}
+//	  total: 2
+//	  gpt-5.4: 1
+//	  gpt-5.3-codex-spark: 1
+type ConcurrencyLimit struct {
+	Total  int
+	Models map[string]int
+}
+
+// UnmarshalYAML handles both int and map forms.
+func (c *ConcurrencyLimit) UnmarshalYAML(value *yaml.Node) error {
+	// Try int first
+	var n int
+	if err := value.Decode(&n); err == nil {
+		c.Total = n
+		return nil
+	}
+	// Try map
+	var m map[string]int
+	if err := value.Decode(&m); err != nil {
+		return err
+	}
+	c.Total = m["total"]
+	c.Models = make(map[string]int)
+	for k, v := range m {
+		if k != "total" {
+			c.Models[k] = v
+		}
+	}
+	return nil
+}
+
 // AgentsConfig is the root of config/agents.yaml.
 //
 //	cfg := agentic.AgentsConfig{Version: 1, Dispatch: agentic.DispatchConfig{DefaultAgent: "claude"}}
 type AgentsConfig struct {
-	Version     int                   `yaml:"version"`
-	Dispatch    DispatchConfig        `yaml:"dispatch"`
-	Concurrency map[string]int        `yaml:"concurrency"`
-	Rates       map[string]RateConfig `yaml:"rates"`
+	Version     int                          `yaml:"version"`
+	Dispatch    DispatchConfig               `yaml:"dispatch"`
+	Concurrency map[string]ConcurrencyLimit  `yaml:"concurrency"`
+	Rates       map[string]RateConfig        `yaml:"rates"`
 }
 
 // loadAgentsConfig reads config/agents.yaml from the code path.
@@ -66,9 +101,9 @@ func (s *PrepSubsystem) loadAgentsConfig() *AgentsConfig {
 			DefaultAgent:    "claude",
 			DefaultTemplate: "coding",
 		},
-		Concurrency: map[string]int{
-			"claude": 1,
-			"gemini": 3,
+		Concurrency: map[string]ConcurrencyLimit{
+			"claude": {Total: 1},
+			"gemini": {Total: 3},
 		},
 	}
 }
@@ -142,6 +177,28 @@ func (s *PrepSubsystem) countRunningByAgent(agent string) int {
 	return count
 }
 
+// countRunningByModel counts running workspaces for a specific agent:model string.
+func (s *PrepSubsystem) countRunningByModel(agent string) int {
+	wsRoot := WorkspaceRoot()
+	old := core.PathGlob(core.JoinPath(wsRoot, "*", "status.json"))
+	deep := core.PathGlob(core.JoinPath(wsRoot, "*", "*", "*", "status.json"))
+
+	count := 0
+	for _, statusPath := range append(old, deep...) {
+		st, err := ReadStatus(core.PathDir(statusPath))
+		if err != nil || st.Status != "running" {
+			continue
+		}
+		if st.Agent != agent {
+			continue
+		}
+		if st.PID > 0 && syscall.Kill(st.PID, 0) == nil {
+			count++
+		}
+	}
+	return count
+}
+
 // baseAgent strips the model variant (gemini:flash → gemini).
 func baseAgent(agent string) string {
 	// codex:gpt-5.3-codex-spark → codex-spark (separate pool)
@@ -151,15 +208,48 @@ func baseAgent(agent string) string {
 	return core.SplitN(agent, ":", 2)[0]
 }
 
-// canDispatchAgent checks if we're under the concurrency limit for a specific agent type.
+// canDispatchAgent checks both pool-level and per-model concurrency limits.
+//
+//	codex: {total: 2, models: {gpt-5.4: 1}} → max 2 codex total, max 1 gpt-5.4
 func (s *PrepSubsystem) canDispatchAgent(agent string) bool {
 	cfg := s.loadAgentsConfig()
 	base := baseAgent(agent)
 	limit, ok := cfg.Concurrency[base]
-	if !ok || limit <= 0 {
+	if !ok || limit.Total <= 0 {
 		return true
 	}
-	return s.countRunningByAgent(base) < limit
+
+	// Check pool total
+	if s.countRunningByAgent(base) >= limit.Total {
+		return false
+	}
+
+	// Check per-model limit if configured
+	if limit.Models != nil {
+		model := modelVariant(agent)
+		if model != "" {
+			if modelLimit, has := limit.Models[model]; has && modelLimit > 0 {
+				if s.countRunningByModel(agent) >= modelLimit {
+					return false
+				}
+			}
+		}
+	}
+
+	return true
+}
+
+// modelVariant extracts the model name from an agent string.
+//
+//	codex:gpt-5.4 → gpt-5.4
+//	codex:gpt-5.3-codex-spark → gpt-5.3-codex-spark
+//	claude → ""
+func modelVariant(agent string) string {
+	parts := core.SplitN(agent, ":", 2)
+	if len(parts) < 2 {
+		return ""
+	}
+	return parts[1]
 }
 
 // drainQueue fills all available concurrency slots from queued workspaces.
diff --git a/pkg/agentic/queue_test.go b/pkg/agentic/queue_test.go
index c3a1645..3f54a37 100644
--- a/pkg/agentic/queue_test.go
+++ b/pkg/agentic/queue_test.go
@@ -26,8 +26,8 @@ func TestDispatchConfig_Good_Defaults(t *testing.T) {
 	cfg := s.loadAgentsConfig()
 	assert.Equal(t, "claude", cfg.Dispatch.DefaultAgent)
 	assert.Equal(t, "coding", cfg.Dispatch.DefaultTemplate)
-	assert.Equal(t, 1, cfg.Concurrency["claude"])
-	assert.Equal(t, 3, cfg.Concurrency["gemini"])
+	assert.Equal(t, 1, cfg.Concurrency["claude"].Total)
+	assert.Equal(t, 3, cfg.Concurrency["gemini"].Total)
 }
 
 func TestCanDispatchAgent_Good_NoConfig(t *testing.T) {