diff --git a/charts/community-components/templates/agent-claude-code-swe.yaml b/charts/community-components/templates/agent-claude-code-swe.yaml
index 8bd6126..4c3b8a9 100644
--- a/charts/community-components/templates/agent-claude-code-swe.yaml
+++ b/charts/community-components/templates/agent-claude-code-swe.yaml
@@ -26,6 +26,11 @@ metadata:
   name: claude-code-swe-agent
   labels:
     {{- include "tools.labels" . | nindent 4 }}
+  annotations:
+    # Tells the Temporal engine's agentWorkflowNameFor (engines/temporal/internal/temporal/workflows/agent_workflow.go)
+    # to drive this Agent via BridgedAgentWorkflow -- it's a pod running the
+    # Claude Code CLI headless, not a declarative Tool-calling planner loop.
+    durable-agents.dev/bridged: "true"
 spec:
   description: >-
     Performs software-engineering work on GitHub end-to-end. Runs the Claude
diff --git a/charts/community-components/templates/agent-opencode-swe.yaml b/charts/community-components/templates/agent-opencode-swe.yaml
index 4e8df71..b24f1c9 100644
--- a/charts/community-components/templates/agent-opencode-swe.yaml
+++ b/charts/community-components/templates/agent-opencode-swe.yaml
@@ -22,6 +22,11 @@ metadata:
   name: opencode-swe-agent
   labels:
     {{- include "tools.labels" . | nindent 4 }}
+  annotations:
+    # Tells the Temporal engine's agentWorkflowNameFor (engines/temporal/internal/temporal/workflows/agent_workflow.go)
+    # to drive this Agent via BridgedAgentWorkflow -- it's a pod running the
+    # opencode CLI headless, not a declarative Tool-calling planner loop.
+    durable-agents.dev/bridged: "true"
 spec:
   description: >-
     Performs software-engineering work on GitHub end-to-end. Runs the
diff --git a/charts/community-components/templates/agent-stub.yaml b/charts/community-components/templates/agent-stub.yaml
index 3ee6b51..d4429d7 100644
--- a/charts/community-components/templates/agent-stub.yaml
+++ b/charts/community-components/templates/agent-stub.yaml
@@ -15,6 +15,14 @@ metadata:
   labels:
     {{- include "tools.labels" . | nindent 4 }}
     e2e: "true"
+  annotations:
+    # Mirrors claude-code-swe-agent/opencode-swe-agent (the pod/NATS-protocol
+    # agents this stands in for): without this, the Temporal engine's
+    # agentWorkflowNameFor (engines/temporal/internal/temporal/workflows/agent_workflow.go)
+    # would route this Agent to the declarative planner loop instead of
+    # BridgedAgentWorkflow, making the stub an unfaithful stand-in once a
+    # cluster routes turns through the Temporal engine.
+    durable-agents.dev/bridged: "true"
 spec:
   description: >-
     E2E test double. Speaks the real NATS agent protocol and returns a canned
diff --git a/e2e/specs/bridged-agent-routing.e2e.ts b/e2e/specs/bridged-agent-routing.e2e.ts
new file mode 100644
index 0000000..f343187
--- /dev/null
+++ b/e2e/specs/bridged-agent-routing.e2e.ts
@@ -0,0 +1,53 @@
+import { describe, expect, it } from "vitest";
+import { requireMinikubeContext } from "../support/guard.js";
+import { kubectlJson } from "../support/k8s.js";
+
+requireMinikubeContext();
+
+/**
+ * The Temporal engine (docs/adr/0036) decides which workflow drives an Agent
+ * purely from a CR annotation -- `durable-agents.dev/bridged: "true"` means
+ * `BridgedAgentWorkflow` (an unmodified upstream pod/CLI agent, speaking the
+ * real NATS protocol); its absence falls through to the declarative
+ * `AgentWorkflow`, an LLM planner that only ever calls `Tool`s named in the
+ * agent's own `toolRefs`.
+ *
+ * `claude-code-swe-agent` and `opencode-swe-agent` are pod-based coding
+ * agents: their `agentPrompt` tells the model to invoke `git`/`gh` as plain
+ * CLI commands, and neither declares any `toolRefs`. Deployed WITHOUT the
+ * annotation, both silently fell through to the declarative loop, which
+ * handed the planner that same prompt but an EMPTY tool list -- so it tried
+ * to call "gh" as a declarative Tool and got refused with "tool not
+ * available to this agent". That is a real incident this asserts against,
+ * not a hypothetical: charts/community-components/templates/
+ * agent-claude-code-swe.yaml and agent-opencode-swe.yaml never set the
+ * annotation until this fix.
+ *
+ * This can't observe an actual `BridgedAgentWorkflow` execution end-to-end
+ * (this minikube profile has no Temporal server deployed alongside the
+ * engine), so it asserts the one thing that IS verifiable here and is
+ * exactly what regressed: the live, cluster-deployed `Agent` CR objects
+ * carry the annotation a Helm chart edit could silently drop again.
+ * `stub-agent` is included because it stands in for claude-code-swe-agent in
+ * happy-path.e2e.ts and must route identically to stay a faithful stand-in.
+ */
+describe("pod-based agents are annotated for BridgedAgentWorkflow routing", () => {
+  const BRIDGED_ANNOTATION = "durable-agents.dev/bridged";
+
+  // opencode-swe-agent is intentionally excluded here: it is disabled in this
+  // suite's deployed values (no built image to enable it with), so no live CR
+  // exists to assert against. Its routing contract is instead pinned
+  // hermetically in engines/temporal/internal/temporal/workflows/
+  // agent_workflow_routing_test.go (TestPodAgentsRouteBridged), alongside
+  // claude-code-swe-agent.
+  const BRIDGED_AGENTS = ["claude-code-swe-agent", "stub-agent"];
+
+  it.each(BRIDGED_AGENTS)("Agent %s declares the bridged annotation", async (agentName) => {
+    const agent = await kubectlJson<{ metadata?: { annotations?: Record<string, string> } }>([
+      "get",
+      "agent",
+      agentName,
+    ]);
+    expect(agent.metadata?.annotations?.[BRIDGED_ANNOTATION]).toBe("true");
+  });
+});
diff --git a/engines/temporal/internal/temporal/workflows/agent_workflow_routing_test.go b/engines/temporal/internal/temporal/workflows/agent_workflow_routing_test.go
new file mode 100644
index 0000000..5e8bb36
--- /dev/null
+++ b/engines/temporal/internal/temporal/workflows/agent_workflow_routing_test.go
@@ -0,0 +1,77 @@
+package workflows
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/require"
+	"k8s.io/apimachinery/pkg/apis/meta/v1/unstructured"
+
+	"github.com/controller-agent/temporal-engine/internal/catalog"
+)
+
+// TestAgentWorkflowNameFor pins agentWorkflowNameFor's routing contract: a
+// regression here (e.g. an accidental case reorder, or a chart authoring gap
+// clearing an annotation) previously misrouted a pod-based coding agent into
+// the declarative planner loop, which then tried to call a native CLI command
+// ("gh") as a declarative Tool and got refused with "tool not available to
+// this agent" -- see the claude-code-swe-agent/opencode-swe-agent incident
+// this test was added for.
+func TestAgentWorkflowNameFor(t *testing.T) {
+	t.Run("bridged annotation routes to BridgedAgentWorkflow", func(t *testing.T) {
+		name := agentWorkflowNameFor(catalog.AgentDescriptor{ID: "claude-code-swe-agent", Bridged: true})
+		require.Equal(t, BridgedAgentWorkflowName, name)
+	})
+
+	t.Run("step-tool annotation routes to PodAgentWorkflow and wins over bridged", func(t *testing.T) {
+		name := agentWorkflowNameFor(catalog.AgentDescriptor{ID: "x", StepToolRef: "some-tool", Bridged: true})
+		require.Equal(t, PodAgentWorkflowName, name)
+	})
+
+	t.Run("neither annotation falls back to the declarative AgentWorkflow", func(t *testing.T) {
+		name := agentWorkflowNameFor(catalog.AgentDescriptor{ID: "x"})
+		require.Equal(t, AgentWorkflowName, name)
+	})
+}
+
+// TestPodAgentsRouteBridged decodes Agent CRs shaped exactly like what
+// charts/community-components/templates/agent-claude-code-swe.yaml and
+// agent-opencode-swe.yaml render (name + the durable-agents.dev/bridged
+// annotation, ADR 0028) and asserts they resolve to BridgedAgentWorkflow, not
+// the declarative loop.
+//
+// This does not render the real Helm templates (that would need a `helm`
+// binary, unavailable in this module's CI job) -- e2e/specs coverage owns
+// asserting the LIVE deployed CR objects actually carry the annotation. This
+// test instead pins the contract those two chart entries must keep meeting:
+// a pod-based coding agent (image-driven, no toolRefs, an agentPrompt telling
+// the model to invoke its CLI's own bash/gh/git directly) MUST declare
+// `durable-agents.dev/bridged: "true"`, or its planner gets no `tools` and any
+// CLI invocation the model narrates gets misread as a declarative tool call.
+func TestPodAgentsRouteBridged(t *testing.T) {
+	for _, agentID := range []string{"claude-code-swe-agent", "opencode-swe-agent"} {
+		t.Run(agentID, func(t *testing.T) {
+			obj := &unstructured.Unstructured{Object: map[string]any{
+				"apiVersion": "core.controller-agent.dev/v1alpha1",
+				"kind":       "Agent",
+				"metadata": map[string]any{
+					"name": agentID,
+					"annotations": map[string]any{
+						"durable-agents.dev/bridged": "true",
+					},
+				},
+				"spec": map[string]any{
+					"description":  "Performs software-engineering work on GitHub end-to-end.",
+					"allowedRoles": []any{"writer"},
+					// No toolRefs: a bridged pod agent's tools are its CLI's own
+					// native built-ins, never declarative Tool CRs.
+				},
+			}}
+
+			descriptor, err := catalog.DecodeAgent(obj)
+			require.NoError(t, err)
+			require.True(t, descriptor.Bridged, "expected %s to decode with Bridged=true", agentID)
+			require.Empty(t, descriptor.ToolRefs)
+			require.Equal(t, BridgedAgentWorkflowName, agentWorkflowNameFor(descriptor))
+		})
+	}
+}