THEDE TECHNOLOGIES
Revised 19 Nov 2026

Contain the agent, sheet 07 of 13

Sandboxing and containment

Isolating what a coding agent can read, write and reach, and guarding the handoff points where your own machine runs what it wrote.

13Tools I checked
3Open gaps
Products exist, defaults are offWhere the market is
19 Nov 2026Last checked
If you do one thing

Treat the line between the sandbox and your machine as a checklist: hooks, settings files, the Docker socket, the keychain. Then check the sandbox is actually on. In Claude Code it starts off.

The toolkit as a map: 4 layers, 13 sheets, 112 tools drawn as filled squares and 46 open gaps drawn as hatched squares INDEX TO THE THIRTEEN SHEETS 112 tools, 46 open gaps, revised 19 Nov 2026 LAYER 1 TRUST THE OUTPUT Layer total 3 sheets 22 tools on the map 11 open gaps 01 Observability and evals promptfoo 1 promptfoo Langfuse 2 Langfuse Arize Phoenix 3 Arize Phoenix Braintrust 4 Braintrust DeepEval 5 DeepEval OpenTelemetry GenAI conventions 6 OTel GenAI Demystifying evals for AI agents 7 Evals guide OpenAI deprecations 8 OpenAI dates Judge calibration 9 Judge kappa Open gap a a Open gap b b Open gap c c Open gap d d 02 The door before the merge GitHub branch protection and CODEOWNERS 1 Branch rules GitHub Copilot code review 2 Copilot review Claude Code Review 3 Claude Review Greptile 4 Greptile CodeRabbit 5 CodeRabbit Socket Firewall 6 Socket Firewall GitHub Agentic Workflows 7 Agentic Workflows Open gap a a Open gap b b Open gap c c Open gap d d 03 Measuring the org DORA, State of AI-assisted Software Development 2025 1 DORA 2025 METR, early-2025 AI and experienced developers 2 METR study DX AI Measurement Framework 3 DX GitClear code quality research 4 GitClear Faros 5 Faros Multitudes 6 Multitudes Open gap a a Open gap b b Open gap c c LAYER 2 CONTAIN THE AGENT 4 sheets, 35 tools, 13 gaps 04 MCP gateways and registries MCP specification, 2026-07-28 1 MCP spec Official MCP Registry 2 MCP Registry Docker MCP Gateway 3 Docker Gateway Obot 4 Obot Microsoft MCP Gateway 5 Microsoft gateway Mintive 6 Mintive Kong AI Gateway 7 Kong Cloudflare MCP server portals 8 Cloudflare portals Amazon Bedrock AgentCore Gateway 9 AgentCore Gateway Open gap a a Open gap b b Open gap c c Open gap d d Open gap e e 05 Identity and secrets Okta Cross App Access 1 Cross App Access Auth0 Token Vault 2 Token Vault Microsoft Entra Agent ID 3 Entra Agent ID Aembit 4 Aembit Teleport 5 Teleport SPIFFE 6 SPIFFE Open gap a a Open gap b b 06 Shadow AI and ghost jobs Nudge Security 1 Nudge Harmonic Security 2 Harmonic Reco 3 Reco Microsoft Purview for AI 4 Purview Healthchecks.io 5 Health checks Cronitor 6 Cronitor Sentry Crons 7 Sentry Crons Open gap a a Open gap b b Open gap c c 07 Sandboxing and containment Claude Code sandbox 1 Claude Code sandbox Codex sandboxing 2 Codex nono 3 nono Docker Sandboxes 4 Docker Sandboxes E2B 5 E2B Hosted sandboxes 6 Hosted sandboxes Apple container 7 Apple container Anthropic Sandbox Runtime 8 Sandbox Runtime Agent hooks 9 Agent hooks Claude Code managed settings 10 Managed settings How we built Claude Code auto mode 11 Auto mode The lethal trifecta 12 Lethal trifecta OWASP LLM06: Excessive Agency 13 OWASP LLM06 Open gap a a Open gap b b Open gap c c LAYER 3 PAY FOR IT Layer total 2 sheets 13 tools on the map 7 open gaps 08 Spend governance Claude plans and pricing 1 Claude plans Claude Code cost controls 2 Claude Code costs GitHub budgets and alerts 3 GitHub budgets LiteLLM 4 LiteLLM Cloudflare AI Gateway 5 Cloudflare AI Gateway Vantage for Anthropic 6 Vantage Open gap a a Open gap b b Open gap c c Open gap d d 09 Routing and model lifecycle Anthropic model deprecations 1 Claude dates OpenRouter Auto Router 2 OpenRouter Auto Microsoft Foundry model router 3 Foundry router Amazon Bedrock intelligent prompt routing 4 Bedrock routing Not Diamond 5 Not Diamond Requesty 6 Requesty RouteLLM 7 RouteLLM Open gap a a Open gap b b Open gap c c LAYER 4 FEED IT, STAFF IT, SELL IT 4 sheets, 42 tools, 15 gaps 10 Context layer and memory Glean 1 Glean Onyx 2 Onyx Credal 3 Credal Dust 4 Dust msgvault 5 msgvault AGENTS.md 6 AGENTS .md Mem0 7 Mem0 Open gap a a Open gap b b Open gap c c Open gap d d Open gap e e 11 Hiring and verification Greenhouse Real Talent 1 Greenhouse Real Talent Pindrop 2 Pindrop Karat 3 Karat CoderPad 4 CoderPad Open gap a a Open gap b b 12 Discovery and selling first User Interviews 1 User Interviews Respondent 2 Respondent Boardy 3 Boardy Happenstance 4 Happen stance Clay 5 Clay Listen Labs 6 Listen Labs Synthetic Users 7 Synthetic Users Synthetic Users: If, When, and How to Use AI-Generated Research 8 NN/g article Open gap a a Open gap b b Open gap c c 13 Agent harnesses Claude Code 1 Claude Code OpenAI Codex 2 Codex Gemini CLI 3 Gemini CLI GitHub Copilot agents 4 Copilot agents Cursor 5 Cursor OpenCode 6 OpenCode Goose 7 Goose Cline 8 Cline Pi, and its fork omp 9 Pi Claude Agent SDK 10 Claude Agent SDK OpenAI Agents SDK 11 OpenAI Agents SDK Google Agent Development Kit 12 Google ADK Microsoft Agent Framework 13 Agent Framework Strands Agents 14 Strands LangGraph 15 LangGraph Effective harnesses for long-running agents 16 Harness guide Paradev, formerly Superset 17 Paradev Orca 18 Orca Herdr 19 Herdr Paseo 20 Paseo Symphony 21 Symphony Stripe Minions 22 Stripe Minions StrongDM's software factory 23 StrongDM factory Open gap a a Open gap b b Open gap c c Open gap d d Open gap e e The toolkit as a map: 4 layers, 13 sheets, 112 tools drawn as filled squares and 46 open gaps drawn as hatched squares INDEX TO THE THIRTEEN SHEETS 112 tools, 46 open gaps, revised 19 Nov 2026 LAYER 1 TRUST THE OUTPUT Layer total 3 sheets 22 tools on the map 11 open gaps 01 Observability and evals promptfoo 1 promptfoo Langfuse 2 Langfuse Arize Phoenix 3 Arize Phoenix Braintrust 4 Braintrust DeepEval 5 DeepEval OpenTelemetry GenAI conventions 6 OTel GenAI Demystifying evals for AI agents 7 Evals guide OpenAI deprecations 8 OpenAI dates Judge calibration 9 Judge kappa Open gap a a Open gap b b Open gap c c Open gap d d 02 The door before the merge GitHub branch protection and CODEOWNERS 1 Branch rules GitHub Copilot code review 2 Copilot review Claude Code Review 3 Claude Review Greptile 4 Greptile CodeRabbit 5 CodeRabbit Socket Firewall 6 Socket Firewall GitHub Agentic Workflows 7 Agentic Workflows Open gap a a Open gap b b Open gap c c Open gap d d 03 Measuring the org DORA, State of AI-assisted Software Development 2025 1 DORA 2025 METR, early-2025 AI and experienced developers 2 METR study DX AI Measurement Framework 3 DX GitClear code quality research 4 GitClear Faros 5 Faros Multitudes 6 Multitudes Open gap a a Open gap b b Open gap c c LAYER 2 CONTAIN THE AGENT 4 sheets, 35 tools, 13 gaps 04 MCP gateways and registries MCP specification, 2026-07-28 1 MCP spec Official MCP Registry 2 MCP Registry Docker MCP Gateway 3 Docker Gateway Obot 4 Obot Microsoft MCP Gateway 5 Microsoft gateway Mintive 6 Mintive Kong AI Gateway 7 Kong Cloudflare MCP server portals 8 Cloudflare portals Amazon Bedrock AgentCore Gateway 9 AgentCore Gateway Open gap a a Open gap b b Open gap c c Open gap d d Open gap e e 05 Identity and secrets Okta Cross App Access 1 Cross App Access Auth0 Token Vault 2 Token Vault Microsoft Entra Agent ID 3 Entra Agent ID Aembit 4 Aembit Teleport 5 Teleport SPIFFE 6 SPIFFE Open gap a a Open gap b b 06 Shadow AI and ghost jobs Nudge Security 1 Nudge Harmonic Security 2 Harmonic Reco 3 Reco Microsoft Purview for AI 4 Purview Healthchecks.io 5 Health checks Cronitor 6 Cronitor Sentry Crons 7 Sentry Crons Open gap a a Open gap b b Open gap c c 07 Sandboxing and containment Claude Code sandbox 1 Claude Code sandbox Codex sandboxing 2 Codex nono 3 nono Docker Sandboxes 4 Docker Sandboxes E2B 5 E2B Hosted sandboxes 6 Hosted sandboxes Apple container 7 Apple container Anthropic Sandbox Runtime 8 Sandbox Runtime Agent hooks 9 Agent hooks Claude Code managed settings 10 Managed settings How we built Claude Code auto mode 11 Auto mode The lethal trifecta 12 Lethal trifecta OWASP LLM06: Excessive Agency 13 OWASP LLM06 Open gap a a Open gap b b Open gap c c LAYER 3 PAY FOR IT Layer total 2 sheets 13 tools on the map 7 open gaps 08 Spend governance Claude plans and pricing 1 Claude plans Claude Code cost controls 2 Claude Code costs GitHub budgets and alerts 3 GitHub budgets LiteLLM 4 LiteLLM Cloudflare AI Gateway 5 Cloudflare AI Gateway Vantage for Anthropic 6 Vantage Open gap a a Open gap b b Open gap c c Open gap d d 09 Routing and model lifecycle Anthropic model deprecations 1 Claude dates OpenRouter Auto Router 2 OpenRouter Auto Microsoft Foundry model router 3 Foundry router Amazon Bedrock intelligent prompt routing 4 Bedrock routing Not Diamond 5 Not Diamond Requesty 6 Requesty RouteLLM 7 RouteLLM Open gap a a Open gap b b Open gap c c LAYER 4 FEED IT, STAFF IT, SELL IT 4 sheets, 42 tools, 15 gaps 10 Context layer and memory Glean 1 Glean Onyx 2 Onyx Credal 3 Credal Dust 4 Dust msgvault 5 msgvault AGENTS.md 6 AGENTS .md Mem0 7 Mem0 Open gap a a Open gap b b Open gap c c Open gap d d Open gap e e 11 Hiring and verification Greenhouse Real Talent 1 Greenhouse Real Talent Pindrop 2 Pindrop Karat 3 Karat CoderPad 4 CoderPad Open gap a a Open gap b b 12 Discovery and selling first User Interviews 1 User Interviews Respondent 2 Respondent Boardy 3 Boardy Happenstance 4 Happen stance Clay 5 Clay Listen Labs 6 Listen Labs Synthetic Users 7 Synthetic Users Synthetic Users: If, When, and How to Use AI-Generated Research 8 NN/g article Open gap a a Open gap b b Open gap c c 13 Agent harnesses Claude Code 1 Claude Code OpenAI Codex 2 Codex Gemini CLI 3 Gemini CLI GitHub Copilot agents 4 Copilot agents Cursor 5 Cursor OpenCode 6 OpenCode Goose 7 Goose Cline 8 Cline Pi, and its fork omp 9 Pi Claude Agent SDK 10 Claude Agent SDK OpenAI Agents SDK 11 OpenAI Agents SDK Google Agent Development Kit 12 Google ADK Microsoft Agent Framework 13 Agent Framework Strands Agents 14 Strands LangGraph 15 LangGraph Effective harnesses for long-running agents 16 Harness guide Paradev, formerly Superset 17 Paradev Orca 18 Orca Herdr 19 Herdr Paseo 20 Paseo Symphony 21 Symphony Stripe Minions 22 Stripe Minions StrongDM's software factory 23 StrongDM factory Open gap a a Open gap b b Open gap c c Open gap d d Open gap e e
Sheet 07: numbers are tools, letters are gapsOne sheet to the view5 tools

← → move to the next sheet over, Esc zooms out.

The problem, in plain words

A coding agent runs real commands on a real computer. A sandbox decides which files it can read, which it can change, and which parts of the internet it can reach. The operating system enforces it, so it holds even when the agent is confidently wrong.

Think of the valet key. It starts the car and won't open the glovebox or the trunk. Some agents hand you the valet key by default. Others leave it in a drawer and give out the full set, and you have to go get the valet key on purpose.

The clever escape isn't through the engine, either. The agent leaves a note on the dashboard (a git hook, a settings file, a script, a plugin) and the next trusted driver follows it without reading. Most of the work here is guarding those handoff points.

Words you'll hear

Sandbox
An enforced boundary around what a process can read, write and reach. The operating system enforces it, not a polite request.
Seatbelt, bubblewrap
The sandboxing tools built into macOS and available on Linux. Most coding agents use one or the other underneath.
microVM
A tiny virtual machine per agent. A stronger wall than a process sandbox, and a heavier one.
Permission prompt
The "allow this command?" box. People approve nearly all of them, which is the problem.
Auto mode
A classifier that approves or blocks the agent's actions instead of asking you.
Phantom token
A stand-in credential the agent can use but never see. The real secret is swapped in outside the sandbox.
Lethal trifecta
Private data, untrusted content and a way to send data out. An agent with all three can be talked into stealing.
Excessive agency
OWASP's name for an agent with more functionality, permission or autonomy than its job needs.

What to understand, and what to decide

Understand

Decide

Who's working on it

Every link opened on 19 Nov 2026. The numbers match the squares on the map.

  1. Claude Code sandbox

    Built in

    Seatbelt on macOS, bubblewrap on Linux and WSL2, and a proxy that lets shell commands reach only the domains you allow.

    The catch. Off by default, and it wraps shell commands only. File tools, MCP servers, hooks and mods run outside it, and ~/.ssh stays readable until you deny it.

  2. Codex sandboxing

    Built in

    Seatbelt on macOS, bubblewrap on Linux, a native sandbox on Windows. On in the default permissions mode.

    The catch. Network access and escalations still come back to an approval. Whoever answers it is your boundary.

  3. nono

    Open source

    A kernel-level sandbox per agent with phantom tokens, so the agent never holds the real credential, plus signed skills, a tamper-evident audit log and rollback.

    The catch. Pre-1.0, so the APIs may still move. Apache-2.0, from the team behind Sigstore.

  4. Docker Sandboxes

    Product

    A microVM per agent, with credentials handled outside the agent and a network policy per sandbox. Runs Claude Code, Codex, Copilot and others.

    The catch. A virtual machine per agent is heavier than a process sandbox. Worth it for untrusted work, overkill for a typo fix.

  5. E2B

    Open source, hosted

    An isolated machine for every agent, in E2B's cloud or your own AWS, GCP or Azure account. The SDK is Apache-2.0.

    The catch. Your code and data run in somebody else's VM unless you bring your own cloud account.

  6. Fly Sprites, Cloudflare Sandboxes, Modal Sandboxes

    Hosted

    Hosted sandboxes for agent code. Sprites persist and checkpoint, Cloudflare offers Linux VMs or isolated Workers, Modal runs gVisor or VMs.

    The catch. Pick by how long the agent's work lives. Throwaway jobs want cheap and disposable. Long sessions want checkpoints.

  7. Apple container

    Open source

    Linux containers, each in its own lightweight VM, on Apple silicon.

    The catch. macOS 26 only. A good local second machine, not a policy layer.

  8. Anthropic Sandbox Runtime

    Open source

    The OS-level sandbox under Claude Code, packaged to wrap any process, local MCP servers included. Apache-2.0.

    The catch. A research preview. Use it to close the MCP gap in the first row, not as your only wall.

  9. Claude Code hooks, Codex hooks, Cursor hooks, Copilot hooks, Gemini CLI hooks

    Built in

    Your own script, run before a tool call, that can refuse it whatever the model decides. In Claude Code a hook's deny holds even with permissions bypassed.

    The catch. Most fail open: a hook that crashes or times out usually lets the call through, and not every harness lets you make it block instead. Codex's own docs call tool hooks a useful guardrail, not a complete enforcement boundary.

  10. Claude Code managed settings, Manage mods

    Built in

    A policy file in a system folder that no user, project or local setting overrides: deny rules, PreToolUse hooks that run ahead of any mod's tool-call hook, and a switch that stops the mods users bring, one Claude wrote mid-session included.

    The catch. A developer who is an administrator on the machine can still edit it. On a machine with no managed settings, no managed MCP file and no Team or Enterprise login, a mod can approve a call your own hook blocked.

  11. How we built Claude Code auto mode

    Reference, Anthropic

    Why permission prompts fail: people approve 93% of them. A two-stage classifier replaces most of the prompts.

    The catch. By Anthropic's own count it misses 17% of overeager actions, on a sample of 52.

  12. The lethal trifecta

    Reference

    Private data, untrusted content and a way to send data out. Give one agent all three and an attacker can steal through it.

    The catch. No model-level fix exists. Break one leg of the triangle for every agent.

  13. OWASP LLM06: Excessive Agency

    Reference, OWASP

    The standard list of what goes wrong when an agent has too much functionality, too many permissions or too much autonomy.

    The catch. A list of failure modes, not a control. Map each one to a setting you actually changed.

Still nobody sells

  1. An audit of which processes on your machine will run files an agent wrote. Every vendor documents the VM. Nobody checks the hooks.
  2. Credential brokering, on by default. The right shape exists in nono and Docker Sandboxes. Most coding agents still inherit the developer's whole keychain.
  3. A canary for every policy hook: a forbidden call, run on a schedule, that has to come back refused. Most hooks fail open, and a broken one looks exactly like a quiet day.

Also on the map