diff --git a/src/gallium/drivers/asahi/agx_batch.c b/src/gallium/drivers/asahi/agx_batch.c
index 75b464328c3..dc48830b9f9 100644
--- a/src/gallium/drivers/asahi/agx_batch.c
+++ b/src/gallium/drivers/asahi/agx_batch.c
@@ -144,6 +144,9 @@ agx_batch_init(struct agx_context *ctx,
    agx_bo_unreference(batch->sampler_heap.bo);
    batch->sampler_heap.bo = NULL;
    batch->sampler_heap.count = 0;
+   batch->vs_scratch = false;
+   batch->fs_scratch = false;
+   batch->cs_scratch = false;
 
    /* We need to emit prim state at the start. Max collides with all. */
    batch->reduced_prim = MESA_PRIM_COUNT;
diff --git a/src/gallium/drivers/asahi/agx_pipe.c b/src/gallium/drivers/asahi/agx_pipe.c
index 3c50baf47ee..10436d73ed1 100644
--- a/src/gallium/drivers/asahi/agx_pipe.c
+++ b/src/gallium/drivers/asahi/agx_pipe.c
@@ -1289,6 +1289,9 @@ agx_flush_batch(struct agx_context *ctx, struct agx_batch *batch)
       return;
    }
 
+   if (batch->cs_scratch)
+      agx_batch_add_bo(batch, ctx->scratch_cs.buf);
+
    assert(batch->initialized);
 
    /* Finalize the encoder */
@@ -1339,6 +1342,11 @@ agx_flush_batch(struct agx_context *ctx, struct agx_batch *batch)
       batch->occlusion_buffer.gpu = 0;
    }
 
+   if (batch->vs_scratch)
+      agx_batch_add_bo(batch, ctx->scratch_vs.buf);
+   if (batch->fs_scratch)
+      agx_batch_add_bo(batch, ctx->scratch_fs.buf);
+
    /* TODO: Linux UAPI submission */
    (void)dev;
    (void)zbias;
@@ -1390,6 +1398,10 @@ agx_destroy_context(struct pipe_context *pctx)
 
    pipe_resource_reference(&ctx->heap, NULL);
 
+   agx_scratch_fini(&ctx->scratch_vs);
+   agx_scratch_fini(&ctx->scratch_fs);
+   agx_scratch_fini(&ctx->scratch_cs);
+
    ralloc_free(ctx);
 }
 
@@ -1507,6 +1519,10 @@ agx_create_context(struct pipe_screen *screen, void *priv, unsigned flags)
    ctx->support_lod_bias = !(flags & PIPE_CONTEXT_NO_LOD_BIAS);
    ctx->robust = (flags & PIPE_CONTEXT_ROBUST_BUFFER_ACCESS);
 
+   agx_scratch_init(agx_device(screen), &ctx->scratch_vs);
+   agx_scratch_init(agx_device(screen), &ctx->scratch_fs);
+   agx_scratch_init(agx_device(screen), &ctx->scratch_cs);
+
    return pctx;
 }
 
diff --git a/src/gallium/drivers/asahi/agx_state.h b/src/gallium/drivers/asahi/agx_state.h
index d686bd95cfb..9ded768d23c 100644
--- a/src/gallium/drivers/asahi/agx_state.h
+++ b/src/gallium/drivers/asahi/agx_state.h
@@ -13,6 +13,7 @@
 #include "asahi/lib/agx_device.h"
 #include "asahi/lib/agx_nir_lower_vbo.h"
 #include "asahi/lib/agx_pack.h"
+#include "asahi/lib/agx_scratch.h"
 #include "asahi/lib/agx_tilebuffer.h"
 #include "asahi/lib/pool.h"
 #include "asahi/lib/shaders/geometry.h"
@@ -381,6 +382,11 @@ struct agx_batch {
 
    /* Actual pointer in a uniform */
    struct agx_bo *geom_params_bo;
+
+   /* Whether each stage uses scratch */
+   bool vs_scratch;
+   bool fs_scratch;
+   bool cs_scratch;
 };
 
 struct agx_zsa {
@@ -633,6 +639,10 @@ struct agx_context {
    uint32_t dummy_syncobj;
    int in_sync_fd;
    uint32_t in_sync_obj;
+
+   struct agx_scratch scratch_vs;
+   struct agx_scratch scratch_fs;
+   struct agx_scratch scratch_cs;
 };
 
 static void