diff --git a/apps/worker/src/env.ts b/apps/worker/src/env.ts
index 8a852ec..54a65d9 100644
--- a/apps/worker/src/env.ts
+++ b/apps/worker/src/env.ts
@@ -38,11 +38,11 @@ const envSchema = z.object({
     .enum(["minimal", "openhands"])
     .optional()
     .default("minimal"),
-  /** OpenHands agent-server URL (local: http://localhost:18000, Cloud: from conversation). */
+  /** OpenHands agent-server URL. Docker: http://agent-server:8001, Local: http://localhost:18000 */
   AGENT_SERVER_URL: z
     .string()
     .optional()
-    .transform((value) => value?.trim() || "http://localhost:18000"),
+    .transform((value) => value?.trim() || "http://agent-server:8001"),
   /** OpenHands session API key for agent-server auth. */
   SESSION_API_KEY: z
     .string()
diff --git a/docker/docker-compose.agent-server.yml b/docker/docker-compose.agent-server.yml
new file mode 100644
index 0000000..67baa32
--- /dev/null
+++ b/docker/docker-compose.agent-server.yml
@@ -0,0 +1,51 @@
+# OpenHands Agent Server for Submatter
+# Deploys on Dokploy (CT 102) alongside the existing stack
+# Worker connects via internal Docker DNS: http://agent-server:8001
+
+version: '3.8'
+
+services:
+  agent-server:
+    image: ghcr.io/open-hands/agent-server:0.19.0  # pin to stable version
+    container_name: echodo-agent-server
+    ports:
+      - "8001:8001"  # agent-server API
+    environment:
+      # Server config
+      - AGENT_SERVER_HOST=0.0.0.0
+      - AGENT_SERVER_PORT=8001
+      - LOCAL_BACKEND_API_KEY=${LOCAL_BACKEND_API_KEY}
+      # LLM config for the agent-server's own operations
+      - LLM_MODEL=${OMNIROUTE_MODEL:-auto/coding}
+      - LLM_API_KEY=${OMNIROUTE_API_KEY}
+      - LLM_BASE_URL=${OMNIROUTE_BASE_URL}
+      # Sandbox config
+      - SANDBOX_TYPE=docker
+      - SANDBOX_DOCKER_IMAGE=ghcr.io/open-hands/sandbox:latest
+      - SANDBOX_TIMEOUT=300
+      # Workspace persistence
+      - WORKSPACE_BASE=/workspace
+    volumes:
+      - agent-server-workspace:/workspace
+      - /var/run/docker.sock:/var/run/docker.sock  # for sandbox containers
+    networks:
+      - echodo-network
+    restart: unless-stopped
+    healthcheck:
+      test: ["CMD", "curl", "-f", "http://localhost:8001/server_info"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 10s
+    deploy:
+      resources:
+        limits:
+          memory: 2G
+          cpus: '2'
+
+volumes:
+  agent-server-workspace:
+
+networks:
+  echodo-network:
+    external: true
diff --git a/docs/AGENT-AUTONOMOUS-CONVOYS.md b/docs/AGENT-AUTONOMOUS-CONVOYS.md
index 4646a46..6a54e16 100644
--- a/docs/AGENT-AUTONOMOUS-CONVOYS.md
+++ b/docs/AGENT-AUTONOMOUS-CONVOYS.md
@@ -435,15 +435,42 @@ By end of Phase 4:
 
 ---
 
+## Phase 5: ACP Harness & Runtime Agnosticism (post-v0.1)
+
+**Goal:** Make submatter runtime-agnostic via Agent Client Protocol (ACP).
+
+**Why:** ACP is an emerging standard that allows OpenHands, Claude Code, Gemini CLI, and other runtimes to be swapped via a single protocol. This prevents vendor lock-in and lets operators choose the best runtime per task.
+
+**Tasks:**
+
+1. Add `HarnessType: "acp"` to shared types
+2. Create `apps/worker/src/executor/acp-executor.ts` — implements ACP client (connects to any ACP-compatible server)
+3. Support ACP discovery via `/.well-known/agent.json` (runtime metadata)
+4. Add `acp` harness type to executor registry
+5. Support sub-agents via ACP Task Tool Set
+
+**ACP-compatible runtimes:**
+- OpenHands (primary)
+- Claude Code (Anthropic)
+- Gemini CLI (Google)
+- Cursor (experimental)
+- Custom (any ACP server)
+
+**This is NOT required for v0.1 testing.** Ships after v0.1 ships and proves out.
+
+---
+
 ## Rollout & Scaling
 
 Once Phase 4 succeeds:
 
 1. **Publish to v0.1** — Mark the orchestrator + worker as production-ready for autonomous convoys
-2. **Scale briefs** — Extend to 5–6 briefs per convoy (architect → implementer → reviewer → ops-copilot → etc.)
-3. **Scale roles** — Add `role-docs-writer`, `role-security-auditor`, `role-performance-reviewer` as needed
-4. **Scale models** — Extend omniroute integration to support Claude/OpenAI/Gemini APIs side-by-side
-5. **Migrate Cursor convoys** — For convoys that have historically run in Cursor, run them via OpenHands instead (backward compat: human-initiated, not automated)
+2. **Deploy OpenHands Agent Server** — Docker container on Dokploy (CT 102), worker connects via internal Docker DNS
+3. **Scale briefs** — Extend to 5–6 briefs per convoy (architect → implementer → reviewer → ops-copilot → etc.)
+4. **Scale roles** — Add `role-docs-writer`, `role-security-auditor`, `role-performance-reviewer` as needed
+5. **Scale models** — Extend omniroute integration to support Claude/OpenAI/Gemini APIs side-by-side
+6. **Migrate Cursor convoys** — For convoys that have historically run in Cursor, run them via OpenHands instead (backward compat: human-initiated, not automated)
+7. **ACP harness** — After v0.1 ships, add ACP support to enable runtime-agnostic agent execution
 
 ---