diff --git a/agents/dev/Dockerfile b/agents/dev/Dockerfile
index cff32e0b11b1b..03a0472f00bef 100644
--- a/agents/dev/Dockerfile
+++ b/agents/dev/Dockerfile
@@ -3,7 +3,7 @@
 # desktop with VNC access, the `desktop` helper and the computer-use skill. This image adds what
 # developing T3 Code needs.
 
-FROM ghcr.io/digdir/agentctl-base:ubuntu-26.04-desktop@sha256:ec5ce4ba36f5f8a71078fb02703d4604d28411fc3a035f2ee2c24d1febb3839e
+FROM ghcr.io/digdir/agentctl-base:ubuntu-26.04-desktop@sha256:762a7fd2fc929cddcd1ee8ab90a203c963518719e904fca7a6802111796e2cc7
 
 # The base ends as the unprivileged Agent user; installing packages needs root again.
 USER root
@@ -73,11 +73,11 @@ RUN case "${TARGETARCH}" in \
     && chmod -R a+rX "${PLAYWRIGHT_BROWSERS_PATH}" \
     && rm -rf "${HOME}" /root/.npm
 
-# agentctl and agentd from digdir-agents main, which the Agent needs until a release ships what it
-# relies on: nested Sandboxes, agentd's /dev/kvm and Podman setup, the SSH OOM policy, and the
-# Agent sandbox section, agentctl skill and flagless logins it gives nested Agents. The host runs
-# the same build, and nested Agents run on this one. agentd downloads its runtime on first use.
-COPY --from=ghcr.io/digdir/agentctl:main@sha256:20419bf3c63d835318d6815949741ea272b4d8eaff49e3ff6c759080858a0f38 /agentctl /agentd /usr/local/bin/
+# agentctl and agentd from 0.1.0-preview.13, the first release with what the Agent relies on:
+# nested Sandboxes, agentd's /dev/kvm and Podman setup, the SSH OOM policy, and the Agent sandbox
+# section, agentctl skill and flagless logins it gives nested Agents. The host needs that release
+# or later, and nested Agents run on this one. agentd downloads its runtime on first use.
+COPY --from=ghcr.io/digdir/agentctl:v0.1.0-preview.13@sha256:5451dd5e912865d7ed4544f0bf2e6d0912cd274b878515224cd05896a286b28f /agentctl /agentd /usr/local/bin/
 RUN agentctl --version
 
 COPY rootfs/ /
diff --git a/agents/dev/README.md b/agents/dev/README.md
index f4ef4151d4dc4..b1012cf902d63 100644
--- a/agents/dev/README.md
+++ b/agents/dev/README.md
@@ -12,10 +12,8 @@ The `demo-video` skill films that desktop into captioned demo videos with fframe
 | default  | Fresh clone of `martinothamar/t3code`, `upstream` added |
 | `nested` | Fresh clone, reduced to fit inside the default Agent    |
 
-The manifest needs an agentctl built from digdir-agents `main`, at least the build the
-Dockerfile copies from the `ghcr.io/digdir/agentctl` image, until a release includes
-`nestedSandboxes`: `make user-install` in a digdir-agents checkout installs one.
-`0.1.0-preview.12` and earlier refuse it.
+The manifest needs agentctl `0.1.0-preview.13` or later; earlier releases refuse its
+`nestedSandboxes`.
 
 ```sh
 cp agents/dev/.env.sample agents/dev/.env   # fill in GITHUB_TOKEN and Git identity
diff --git a/agents/dev/agent.nested.yaml b/agents/dev/agent.nested.yaml
index b072450282915..7730f4799c2cd 100644
--- a/agents/dev/agent.nested.yaml
+++ b/agents/dev/agent.nested.yaml
@@ -21,7 +21,3 @@ spec:
         mode: direct
       # Nested Agents of its own would need virtualization two levels down.
       nestedSandboxes: null
-    mounts:
-      - type: tmpfs
-        target: /tmp
-        capacity: "2Gi"
diff --git a/agents/dev/agent.yaml b/agents/dev/agent.yaml
index 7882bc1579c2e..155cbdba79ebe 100644
--- a/agents/dev/agent.yaml
+++ b/agents/dev/agent.yaml
@@ -19,10 +19,6 @@ spec:
       # leaves the Agent without them, as its NestedSandboxes condition says.
       nestedSandboxes: preferred
     initSystem: image
-    mounts:
-      - type: tmpfs
-        target: /tmp
-        capacity: "4Gi"
   home:
     source: home
   instructions: