Agenta-AI · mmabrouk · Jun 20, 2026
diff --git a/.github/workflows/12-check-unit-tests.yml b/.github/workflows/12-check-unit-tests.yml
@@ -301,3 +301,74 @@ jobs:
           files: services/oss/tests/results/junit.xml
           check_name: Application services Unit Test Results
           comment_mode: off
+
+  run-services-node-unit-tests:
+    # The agent runner (services/agent) is a standalone Node/pnpm package, not part of the
+    # Python services suite above. It runs its own vitest unit tests plus a tsc typecheck gate.
+    # No "has_tests" guard on purpose: this suite is established, so a missing/empty suite must
+    # FAIL the job (vitest exits non-zero on no test files), not silently skip it.
+    if: |
+      github.event_name == 'workflow_dispatch' ||
+      !github.event.pull_request.draft
+    runs-on: ubuntu-latest
+    permissions:
+      checks: write
+      pull-requests: write
+      contents: read
+    env:
+      AGENTA_LICENSE: oss
+    steps:
+      - uses: actions/checkout@v6
+
+      - name: Skip when package selection excludes services
+        if: github.event_name == 'workflow_dispatch' && !contains(fromJSON('["all","services-only"]'), inputs.packages)
+        run: exit 0
+
+      - name: Set up Node.js
+        if: github.event_name != 'workflow_dispatch' || contains(fromJSON('["all","services-only"]'), inputs.packages)
+        uses: actions/setup-node@v4
+        with:
+          node-version: '24'
+
+      - name: Enable Corepack
+        if: github.event_name != 'workflow_dispatch' || contains(fromJSON('["all","services-only"]'), inputs.packages)
+        run: corepack enable
+
+      - name: Cache pnpm store
+        if: github.event_name != 'workflow_dispatch' || contains(fromJSON('["all","services-only"]'), inputs.packages)
+        uses: actions/cache@v4
+        with:
+          path: ~/.pnpm-store
+          key: ${{ runner.os }}-services-agent-pnpm-${{ hashFiles('services/agent/pnpm-lock.yaml') }}
+          restore-keys: |
+            ${{ runner.os }}-services-agent-pnpm-
+
+      - name: Set up pnpm store
+        if: github.event_name != 'workflow_dispatch' || contains(fromJSON('["all","services-only"]'), inputs.packages)
+        working-directory: services/agent
+        run: pnpm config set store-dir ~/.pnpm-store
+
+      - name: Install dependencies
+        if: github.event_name != 'workflow_dispatch' || contains(fromJSON('["all","services-only"]'), inputs.packages)
+        working-directory: services/agent
+        run: pnpm install --frozen-lockfile
+
+      - name: Typecheck (tsc --noEmit, src + tests + config)
+        if: github.event_name != 'workflow_dispatch' || contains(fromJSON('["all","services-only"]'), inputs.packages)
+        working-directory: services/agent
+        run: pnpm run typecheck
+
+      # The code-tool unit test spawns python3 and node end-to-end; both are preinstalled on
+      # ubuntu runners (node is also set up above), so no setup-python step is needed.
+      - name: Run agent runner unit tests
+        if: github.event_name != 'workflow_dispatch' || contains(fromJSON('["all","services-only"]'), inputs.packages)
+        working-directory: services/agent
+        run: pnpm run test:unit
+
+      - name: Publish agent runner unit test results
+        if: always() && (github.event_name != 'workflow_dispatch' || contains(fromJSON('["all","services-only"]'), inputs.packages))
+        uses: EnricoMi/publish-unit-test-result-action@v2
+        with:
+          files: services/agent/test-results/junit.xml
+          check_name: Agent Runner Unit Test Results
+          comment_mode: off
diff --git a/.github/workflows/42-railway-build.yml b/.github/workflows/42-railway-build.yml
@@ -109,6 +109,7 @@ jobs:
           - agenta-api
           - agenta-web
           - agenta-services
+          - agenta-sandbox-agent
         arch:
           - amd64
           - arm64
@@ -127,6 +128,9 @@ jobs:
           - image_name: agenta-services
             context: .
             dockerfile: services/oss/docker/Dockerfile.gh
+          - image_name: agenta-sandbox-agent
+            context: services/agent
+            dockerfile: services/agent/docker/Dockerfile
           # Per-arch config (runner + platform string)
           - arch: amd64
             runner: ubuntu-24.04
@@ -215,6 +219,7 @@ jobs:
           - agenta-api
           - agenta-web
           - agenta-services
+          - agenta-sandbox-agent
     steps:
       - name: Log in to GHCR
         uses: docker/login-action@v3

diff --git a/docs/docs/self-host/guides/08-custom-agent-runner-images.mdx b/docs/docs/self-host/guides/08-custom-agent-runner-images.mdx
@@ -0,0 +1,52 @@
+---
+title: Custom Agent Runner Images
+sidebar_label: Custom Runner Images
+description: Build and configure custom sandbox-agent runner images for self-hosted Agenta
+---
+
+The default runner image is:
+
+```bash
+ghcr.io/agenta-ai/agenta-sandbox-agent:latest
+```
+
+Build a custom image when you need extra system packages, custom certificates, or a pinned
+runner build.
+
+## Build
+
+From the repository root:
+
+```bash
+docker build \
+  -f services/agent/docker/Dockerfile \
+  -t ghcr.io/<org>/agenta-sandbox-agent:<tag> \
+  services/agent
+```
+
+The published image does not bundle Claude Code. If a harness needs Claude Code, the runner
+installs it from Anthropic at runtime.
+
+## Configure Compose
+
+```bash
+AGENTA_SANDBOX_AGENT_IMAGE_NAME=<org>/agenta-sandbox-agent
+AGENTA_SANDBOX_AGENT_IMAGE_TAG=<tag>
+```
+
+## Configure Helm
+
+```yaml
+agentRunner:
+  image:
+    repository: ghcr.io/<org>/agenta-sandbox-agent
+    tag: <tag>
+```
+
+## Configure Railway
+
+```bash
+export AGENTA_SANDBOX_AGENT_IMAGE="ghcr.io/<org>/agenta-sandbox-agent:<tag>"
+```
+
+Then run the normal Railway deploy flow.
diff --git a/hosting/docker-compose/ee/docker-compose.gh.yml b/hosting/docker-compose/ee/docker-compose.gh.yml
@@ -291,11 +291,17 @@ services:
         environment:
             - SCRIPT_NAME=/services
             - DOCKER_NETWORK_MODE=${DOCKER_NETWORK_MODE:-bridge}
+            - AGENTA_AGENT_RUNNER_URL=${AGENTA_AGENT_RUNNER_URL:-http://sandbox-agent:8765}
+            - AGENTA_AGENT_ENABLE_MCP=${AGENTA_AGENT_ENABLE_MCP:-false}
         # === NETWORK ============================================== #
         networks:
             - agenta-ee-gh-network
         extra_hosts:
             - "host.docker.internal:host-gateway"
+        # === ORCHESTRATION ======================================== #
+        depends_on:
+            sandbox-agent:
+                condition: service_healthy
         # === LABELS =============================================== #
         labels:
             - "traefik.http.routers.services.rule=PathPrefix(`/services/`)"
@@ -308,6 +314,36 @@ services:
         # === LIFECYCLE ============================================ #
         restart: always
 
+    sandbox-agent:
+        # === IMAGE ================================================ #
+        image: ghcr.io/agenta-ai/${AGENTA_SANDBOX_AGENT_IMAGE_NAME:-internal-ee-agenta-sandbox-agent}:${AGENTA_SANDBOX_AGENT_IMAGE_TAG:-latest}
+        # === CONFIGURATION ======================================== #
+        environment:
+            PORT: "8765"
+            AGENTA_HOST: ${AGENTA_HOST:-}
+            AGENTA_API_KEY: ${AGENTA_API_KEY:-}
+            SANDBOX_AGENT_PROVIDER: ${SANDBOX_AGENT_PROVIDER:-local}
+            SANDBOX_AGENT_DAYTONA_API_KEY: ${SANDBOX_AGENT_DAYTONA_API_KEY:-}
+            SANDBOX_AGENT_DAYTONA_API_URL: ${SANDBOX_AGENT_DAYTONA_API_URL:-}
+            SANDBOX_AGENT_DAYTONA_TARGET: ${SANDBOX_AGENT_DAYTONA_TARGET:-}
+            SANDBOX_AGENT_DAYTONA_SNAPSHOT: ${SANDBOX_AGENT_DAYTONA_SNAPSHOT:-}
+            SANDBOX_AGENT_DAYTONA_IMAGE: ${SANDBOX_AGENT_DAYTONA_IMAGE:-}
+            SANDBOX_AGENT_DAYTONA_INSTALL_PI: ${SANDBOX_AGENT_DAYTONA_INSTALL_PI:-false}
+        # === NETWORK ============================================== #
+        networks:
+            - agenta-ee-gh-network
+        # === LABELS =============================================== #
+        labels:
+            - "traefik.enable=false"
+        # === LIFECYCLE ============================================ #
+        restart: always
+        healthcheck:
+            test: ["CMD", "node", "-e", "fetch('http://127.0.0.1:8765/health').then(r=>process.exit(r.ok?0:1)).catch(()=>process.exit(1))"]
+            interval: 10s
+            timeout: 5s
+            retries: 12
+            start_period: 20s
+
     postgres:
         # === IMAGE ================================================ #
         image: postgres:17

diff --git a/hosting/docker-compose/ee/env.ee.gh.example b/hosting/docker-compose/ee/env.ee.gh.example
@@ -114,12 +114,18 @@ AGENTA_CRYPT_KEY=replace-me
 # CRISP_WEBSITE_ID=
 
 # ================================================================== #
-# daytona
-# ================================================================== #
-# DAYTONA_API_KEY=
-# DAYTONA_API_URL=https://app.daytona.io/api
-# DAYTONA_SNAPSHOT=daytona-small
-# DAYTONA_TARGET=eu
+# sandbox-agent
+# ================================================================== #
+# AGENTA_AGENT_RUNNER_URL=http://sandbox-agent:8765
+# AGENTA_AGENT_ENABLE_MCP=false
+# AGENTA_SANDBOX_AGENT_IMAGE_NAME=internal-ee-agenta-sandbox-agent
+# AGENTA_SANDBOX_AGENT_IMAGE_TAG=latest
+# SANDBOX_AGENT_PROVIDER=local
+# SANDBOX_AGENT_DAYTONA_API_KEY=
+# SANDBOX_AGENT_DAYTONA_API_URL=https://app.daytona.io/api
+# SANDBOX_AGENT_DAYTONA_TARGET=eu
+# SANDBOX_AGENT_DAYTONA_SNAPSHOT=
+# SANDBOX_AGENT_DAYTONA_IMAGE=
 
 # ================================================================== #
 # docker

diff --git a/hosting/docker-compose/oss/docker-compose.gh.yml b/hosting/docker-compose/oss/docker-compose.gh.yml
@@ -318,11 +318,17 @@ services:
         environment:
             - SCRIPT_NAME=/services
             - DOCKER_NETWORK_MODE=${DOCKER_NETWORK_MODE:-bridge}
+            - AGENTA_AGENT_RUNNER_URL=${AGENTA_AGENT_RUNNER_URL:-http://sandbox-agent:8765}
+            - AGENTA_AGENT_ENABLE_MCP=${AGENTA_AGENT_ENABLE_MCP:-false}
         # === NETWORK ============================================== #
         networks:
             - agenta-oss-gh-network
         extra_hosts:
             - "host.docker.internal:host-gateway"
+        # === ORCHESTRATION ======================================== #
+        depends_on:
+            sandbox-agent:
+                condition: service_healthy
         # === LABELS =============================================== #
         labels:
             - "traefik.http.routers.services.rule=PathPrefix(`/services/`)"
@@ -335,6 +341,36 @@ services:
         # === LIFECYCLE ============================================ #
         restart: always
 
+    sandbox-agent:
+        # === IMAGE ================================================ #
+        image: ghcr.io/agenta-ai/${AGENTA_SANDBOX_AGENT_IMAGE_NAME:-agenta-sandbox-agent}:${AGENTA_SANDBOX_AGENT_IMAGE_TAG:-latest}
+        # === CONFIGURATION ======================================== #
+        environment:
+            PORT: "8765"
+            AGENTA_HOST: ${AGENTA_HOST:-}
+            AGENTA_API_KEY: ${AGENTA_API_KEY:-}
+            SANDBOX_AGENT_PROVIDER: ${SANDBOX_AGENT_PROVIDER:-local}
+            SANDBOX_AGENT_DAYTONA_API_KEY: ${SANDBOX_AGENT_DAYTONA_API_KEY:-}
+            SANDBOX_AGENT_DAYTONA_API_URL: ${SANDBOX_AGENT_DAYTONA_API_URL:-}
+            SANDBOX_AGENT_DAYTONA_TARGET: ${SANDBOX_AGENT_DAYTONA_TARGET:-}
+            SANDBOX_AGENT_DAYTONA_SNAPSHOT: ${SANDBOX_AGENT_DAYTONA_SNAPSHOT:-}
+            SANDBOX_AGENT_DAYTONA_IMAGE: ${SANDBOX_AGENT_DAYTONA_IMAGE:-}
+            SANDBOX_AGENT_DAYTONA_INSTALL_PI: ${SANDBOX_AGENT_DAYTONA_INSTALL_PI:-false}
+        # === NETWORK ============================================== #
+        networks:
+            - agenta-oss-gh-network
+        # === LABELS =============================================== #
+        labels:
+            - "traefik.enable=false"
+        # === LIFECYCLE ============================================ #
+        restart: always
+        healthcheck:
+            test: ["CMD", "node", "-e", "fetch('http://127.0.0.1:8765/health').then(r=>process.exit(r.ok?0:1)).catch(()=>process.exit(1))"]
+            interval: 10s
+            timeout: 5s
+            retries: 12
+            start_period: 20s
+
     postgres:
         # === IMAGE ================================================ #
         image: postgres:17

diff --git a/hosting/docker-compose/oss/env.oss.gh.example b/hosting/docker-compose/oss/env.oss.gh.example
@@ -114,12 +114,18 @@ AGENTA_CRYPT_KEY=replace-me
 # CRISP_WEBSITE_ID=
 
 # ================================================================== #
-# daytona
-# ================================================================== #
-# DAYTONA_API_KEY=
-# DAYTONA_API_URL=https://app.daytona.io/api
-# DAYTONA_SNAPSHOT=daytona-small
-# DAYTONA_TARGET=eu
+# sandbox-agent
+# ================================================================== #
+# AGENTA_AGENT_RUNNER_URL=http://sandbox-agent:8765
+# AGENTA_AGENT_ENABLE_MCP=false
+# AGENTA_SANDBOX_AGENT_IMAGE_NAME=agenta-sandbox-agent
+# AGENTA_SANDBOX_AGENT_IMAGE_TAG=latest
+# SANDBOX_AGENT_PROVIDER=local
+# SANDBOX_AGENT_DAYTONA_API_KEY=
+# SANDBOX_AGENT_DAYTONA_API_URL=https://app.daytona.io/api
+# SANDBOX_AGENT_DAYTONA_TARGET=eu
+# SANDBOX_AGENT_DAYTONA_SNAPSHOT=
+# SANDBOX_AGENT_DAYTONA_IMAGE=
 
 # ================================================================== #
 # docker

diff --git a/services/agent/sandbox-images/daytona/README.md b/services/agent/sandbox-images/daytona/README.md
@@ -0,0 +1,25 @@
+# Daytona Sandbox Snapshot
+
+This folder contains the supported self-host recipe for building a Daytona snapshot for the
+Agenta `sandbox-agent` runner path.
+
+We ship the recipe, not a built snapshot. The operator runs it in their own Daytona account:
+
+```bash
+DAYTONA_API_KEY=... DAYTONA_TARGET=eu uv run build_snapshot.py --force
+```
+
+Configure the runner service with:
+
+```bash
+SANDBOX_AGENT_PROVIDER=daytona
+SANDBOX_AGENT_DAYTONA_SNAPSHOT=agenta-sandbox-pi
+SANDBOX_AGENT_DAYTONA_INSTALL_PI=false
+```
+
+The recipe currently bases on the upstream full sandbox-agent image and adds the Pi CLI.
+That image includes Claude Code. We do not distribute the resulting snapshot. Cloud builds its
+own internal snapshot; self-hosters build their own.
+
+Keep credentials out of the image and snapshot. Provider keys and self-managed login paths are
+runtime concerns.