--- type: "Guide" title: "Trapped in Your Own Stack? The 5-Question Check for Managed Agent Models — plus Worktree Clones as a Practical Twin" description: "The 2026 5-question check for managed agent models: state, sandbox, loop, residency, ZDR — plus Git worktrees as a practical twin for parallel agent sessions." resource: "https://www.contextstudios.ai/guides/managed-agent-modelle-5-fragen-check-2026" language: "en" generated: by: "process:contextstudios-md/1" at: "2026-09-26T11:47:24.215Z" status: "stable" --- # Trapped in Your Own Stack? The 5-Question Check for Managed Agent Models — plus Worktree Clones as a Practical Twin Managed agent models like Claude Code, OpenAI Codex and Gemini CLI ship complete agent loops with tool use, session state and parallel subagents — but each vendor defines state, sandbox and retention differently. The 5-question check (state surviving, sandbox type, loop-as-product, data residency, ZDR) quickly separates the fitting models from the rest. For parallel agent sessions, Git worktrees are a good fit: they share the object repository, so they are faster and lighter to create than a second clone. ## The five questions 1. **Question 1 — State surviving** — Does the working state survive a restart? The agent loop writes progress (messages, tool results, task status) into a persistent session state. Test --continue / --resume after a process restart, and check whether the state lives in the project folder (JSONL) or only in the process. Specialization: Session state after restart Price range: With --resume a session can be continued with its history and task state — check after the restart that everything is actually there. 2. **Question 2 — Server-side vs. local sandbox** — Where do shell commands run? Two patterns: local (Claude Code, Codex CLI, OpenCode: process starts next to the repo) or remote (hosted code_interpreter/container tools). Local means zero network latency per tool call; remote means shared, reproducible environments. Specialization: Latency per tool call Price range: Local execution avoids the per-call network overhead; with many small tool calls the overall run is considerably shorter. 3. **Question 3 — Loop-as-product** — Do I get the finished loop or only the model? Managed agent = model + orchestration + tools + sessions from one hand. Model API only = build your own loop (tool calls, max_tokens chains, retry logic). Specialization: Orchestration scope Price range: For multi-step tasks (plan → edit → test → fix) take the finished loop; the three systems above are designed for that. 4. **Question 4 — Data residency** — In which region does the provider process data? Depending on the vendor, region or data residency is selected in the workspace/project configuration or via the cloud provider (region of the cloud deployment). For DACH customers: choose an EU region when latency and EU data location both count. Check availability in the vendor documentation. Specialization: Processing location Price range: A fixed region in the workspace/project or cloud configuration gives you a chain that is documentable in the DPA. 5. **Question 5 — ZDR (Zero Data Retention)** — Is Zero Data Retention available, and under which conditions? ZDR means prompts and outputs are not stored beyond the technically required time windows after the answer. Whether and how ZDR is available (agreement, organization or project setting) is up to each vendor — check the vendor documentation. Specialization: Retention of prompts and outputs Price range: With ZDR: shorter retention, plannable deletion; for internal tools usually the cleaner variant. ZDR plus --resume works while the session is open — long pauses can trim the state. ## How to proceed - Check in this order: resumable sessions, a sandbox type matching the task, a finished loop for multi-step work, control over region or data residency and optionally ZDR — then the stack is a tool, not a cage. - Create the worktree twin: git worktree add ../twin creates a second working directory of the same repository. Because worktrees share the object repository, this is faster and lighter than a second clone. - Work in parallel: in both twins, agent A refactors /api while agent B writes tests. Return with git worktree remove ../twin — the original stays untouched. - Know the limits: a branch can only be checked out in one worktree at a time, and untracked dependencies must be installed per worktree. - Stances: Empowerment — five questions, one terminal, reusable every week without vendor briefings. Pro-local — an EU region or EU data residency plus ZDR, where the vendor offers them, is a minimally invasive, documentable configuration, especially for DACH mid-market companies. ## Frequently asked questions ### Why Git worktrees instead of a second clone? All worktrees of a repository share the same object repository; git worktree add only creates another working directory with its own branch. That is faster and saves disk space compared with a second full clone — and two agents can work in parallel without overwriting each other's files. ### What should I watch out for with worktrees? A branch can only be checked out in one worktree at a time. Untracked dependencies (e.g. node_modules) and local config files are not shared and must be set up per worktree. Clean up with git worktree remove; git worktree prune removes stale entries. ### Does ZDR affect session resume? ZDR plus --resume works as long as the session is still open. Long pauses can trim the state — if you need guaranteed resumption, check how the retention window interacts with your own turn cadence. ### When is a plain model API enough instead of a managed agent? For pure text classification a full agent loop is overkill — a single-call API is cheaper. Take the finished loop for multi-step tasks (plan → edit → test → fix), which Claude Code, OpenAI Agents SDK and Gemini CLI are optimized for. ### Which sandbox type fits many small tool calls? With many small tool calls (fmt, test, grep), a remote sandbox adds a network round trip per call, which adds up noticeably over a long run. Local execution (process next to the repo) avoids that overhead; a remote sandbox in turn offers shared, reproducible and isolated environments.