diff --git a/agentex/database/migrations/alembic/versions/2026_08_03_1500_add_task_states_771e95623724.py b/agentex/database/migrations/alembic/versions/2026_08_03_1500_add_task_states_771e95623724.py
new file mode 100644
index 00000000..48ea95e1
--- /dev/null
+++ b/agentex/database/migrations/alembic/versions/2026_08_03_1500_add_task_states_771e95623724.py
@@ -0,0 +1,73 @@
+"""add task_states
+
+Revision ID: 771e95623724
+Revises: a1b2c3d4e5f6
+Create Date: 2026-08-03 15:00:00.000000
+
+Creates the task_states table: the optional PostgreSQL backend for task state
+(selected per deployment via TASK_STATE_STORAGE_PHASE; MongoDB remains the
+default). Schema-only on a brand-new table, so creation is instant and holds
+no lock against live traffic; nothing reads or writes the table until the
+Postgres task-state repository lands.
+
+The (task_id, agent_id) index is deliberately absent: its shape is the open
+write-semantics decision (a unique constraint backing an atomic upsert, or a
+plain compound index mirroring MongoDB) and it ships with that decision.
+"""
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects.postgresql import JSONB
+
+# revision identifiers, used by Alembic.
+revision: str = "771e95623724"
+down_revision: str | None = "a1b2c3d4e5f6"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "task_states",
+        sa.Column("id", sa.String(), nullable=False),
+        sa.Column("task_id", sa.String(), nullable=False),
+        sa.Column("agent_id", sa.String(), nullable=False),
+        sa.Column("state", JSONB(), nullable=False),
+        sa.Column(
+            "created_at",
+            sa.DateTime(timezone=True),
+            server_default=sa.text("now()"),
+            nullable=False,
+        ),
+        sa.Column(
+            "updated_at",
+            sa.DateTime(timezone=True),
+            server_default=sa.text("now()"),
+            nullable=False,
+        ),
+        # Both FKs are bare (no ON DELETE action), deliberately: MongoDB has no
+        # cascades, so state deletion is application-driven on both backends,
+        # and the FK alone prevents orphaned rows. This matches the sibling
+        # child tables of tasks (task_agents, events), whose bare FKs already
+        # reject the existing DELETE /tasks path for any task with children;
+        # task_states behaves identically rather than cascading on one backend
+        # only.
+        sa.ForeignKeyConstraint(["task_id"], ["tasks.id"]),
+        sa.ForeignKeyConstraint(["agent_id"], ["agents.id"]),
+        sa.PrimaryKeyConstraint("id"),
+    )
+    # The index targets the table created in this same migration, so it holds
+    # no write-blocking lock against live traffic (the table has no rows yet).
+    op.create_index(
+        "ix_task_states_agent_id",
+        "task_states",
+        ["agent_id"],
+        unique=False,
+    )
+
+
+def downgrade() -> None:
+    op.drop_index("ix_task_states_agent_id", table_name="task_states")
+    op.drop_table("task_states")
diff --git a/agentex/src/adapters/orm.py b/agentex/src/adapters/orm.py
index e5f7b139..95a95963 100644
--- a/agentex/src/adapters/orm.py
+++ b/agentex/src/adapters/orm.py
@@ -147,6 +147,49 @@ class AgentTaskTrackerORM(BaseORM):
     )
 
 
+class TaskStateORM(BaseORM):
+    """Task state on PostgreSQL: one row per (task, agent) pair.
+
+    Optional storage backend for task state (selected per deployment via
+    TASK_STATE_STORAGE_PHASE); MongoDB remains the default. Nothing reads or
+    writes this table until the Postgres task-state repository lands.
+    """
+
+    __tablename__ = "task_states"
+
+    id = Column(String, primary_key=True, default=orm_id)
+    # Both FKs are bare (no ON DELETE action), deliberately: MongoDB has no
+    # cascades, so state deletion is application-driven on both backends, and
+    # the FK alone prevents orphaned rows. This matches the sibling child
+    # tables of tasks (task_agents, events), whose bare FKs already reject the
+    # existing DELETE /tasks path for any task with children; task_states
+    # behaves identically rather than cascading on one backend only.
+    task_id = Column(String, ForeignKey("tasks.id"), nullable=False)
+    agent_id = Column(String, ForeignKey("agents.id"), nullable=False)
+    state = Column(JSONB, nullable=False)
+    # NOT NULL with a server default: the repository must OMIT unset (None)
+    # timestamps when constructing rows. SQLAlchemy renders an explicitly
+    # assigned None as a literal NULL, which violates the constraint instead
+    # of falling back to the default (StateEntity defaults these to None).
+    created_at = Column(
+        DateTime(timezone=True), server_default=func.now(), nullable=False
+    )
+    updated_at = Column(
+        DateTime(timezone=True),
+        server_default=func.now(),
+        onupdate=func.now(),
+        nullable=False,
+    )
+
+    __table_args__ = (
+        # The (task_id, agent_id) index is deliberately absent: its shape is
+        # the open write-semantics decision (a unique constraint backing an
+        # atomic upsert, or a plain compound index mirroring MongoDB) and it
+        # ships with the repository that reads this table.
+        Index("ix_task_states_agent_id", "agent_id"),
+    )
+
+
 class SpanORM(BaseORM):
     __tablename__ = "spans"
     id = Column(String, primary_key=True, default=orm_id)  # Using UUIDs for IDs