Skip to content
The AI Agent Factory

Appendix: reproducing the observations

Approved by
Panaversity ·
13 min read

The commands

Run these commands in the Bash tool of a cloud Claude session. Print only the variables listed. Never print values whose names contain TOKEN, KEY or SECRET, because those are credentials. The run date for this section was November 9, 2026, at 02:45 UTC. Version strings and variable names are a snapshot from that date, and they will change.

# 1. Init process and VM uptime
ps -o pid,etime,args -p 1
# 2. Process ancestry of the shell
p=$$; while [ "$p" -gt 1 ]; do ps -o pid=,ppid=,args= -p $p | cut -c1-160; p=$(ps -o ppid= -p $p | tr -d ' '); done
# 3. Claude Code binary and version
which claude; readlink -f "$(which claude)"; claude --version
# 4. Selected environment variables
env | grep -E '^(CLAUDE_CODE_(REMOTE|ENTRYPOINT|SESSION_ORIGIN|VERSION|EXECPATH)|ANTHROPIC_BASE_URL)='
# 5. Working transcript location
ls ~/.claude/projects

Here is what each command does, in plain words:

  1. Shows the first program the computer started (PID 1) and how long it has been running (its uptime).
  2. Follows the chain of programs that started the shell, one parent at a time (its process ancestry).
  3. Finds the claude program, follows any link to the real file, and prints its version.
  4. Prints only the selected environment variables.
  5. Lists the folder where Claude Code keeps its working transcripts.

What each step showed

StepSignal seen on November 9What it supportsLimit
1/process_api --firecracker-init. Uptime about 19 minutesA Firecracker microVM. Inferred: started that morningOne VM at one time. Uptime may not count time spent paused
2The shell is a child of /opt/claude-code/bin/claude, launched by environment-manager task-run --session-mode resumeThe harness is Claude Code, started for each sessionIt does not show what other surfaces use
3The claude command on the PATH reports 2.1.319A Claude Code install is presentIt is the same file as the running program, /opt/claude-code/bin/claude (checked with readlink -f)
4REMOTE=true, ENTRYPOINT=remote_cowork, SESSION_ORIGIN=claude_ai_chat, VERSION=2.1.42, API base api.anthropic.comA remote session started from chat. Model requests go to the configured endpointVariable names may change between releases
5A -home-claude project folderThe transcript is kept in the VMIt is not the durable record

The two VMs, side by side

On November 16, 2026, Panaversity ran a second inspection, in a Claude conversation linked to a Mac with one connected folder. The cloud workspace uses the Bash tool. The desktop workspace uses the device_bash tool, which works only once a folder is connected. Ask Claude to run the same commands with each tool and show you the output.

# Identity, operating system and kernel
whoami; uname -a; cat /etc/os-release | head -3
# CPU, memory and disk
nproc; free -h; df -h $HOME
# Connected folders (desktop workspace only)
ls $HOME/mnt/
# The first running processes
ps aux | head -12
# Common tools
which python3 node git ffmpeg pandoc soffice
SignalCloud workspaceDesktop workspace
User and hostroot on host vmA per-session user, rcw- and the session id, on host claude
Operating systemUbuntu 24.04, x86_64, kernel 6.18 built for FirecrackerUbuntu 22.04.5, arm64 (aarch64), kernel 6.8
CPU, memory, free disk2, 7.8 GiB, about 31 GB4, 3.8 GiB, about 8 GB
First process/process_api --firecracker-init (seen on November 9)bwrap --new-session --unshare-net, with socat relays to a local proxy
Tools foundPython, Node, Git, FFmpeg, LibreOffice and Pandoc, among othersPython 3.10, Node, Git, FFmpeg, Pandoc and LibreOffice
Network testpypi.org answered with HTTP 200pypi.org answered with HTTP 200. A news site did not answer, under the Package managers only setting

Three more signals came from the desktop workspace's own tools on the same day. Staging allowed at most 50 files, 400 MB per file and 500 MB per call. A commit was refused when the file on the computer had changed since it was staged. Deleting in the connected folder failed with "Operation not permitted" until deletion was approved for that folder, which also blocked git until then.

SignalWhat it supportsLimit
Different OS, architecture and sizeTwo separate machines with no shared diskOne account, one Mac, one day
bwrap with --unshare-net as the first processEach session's commands run in their own sandbox inside the local VM, with network only through a proxyThe VM's own design is described only by Anthropic's overview
Staging and commit limitsFiles cross between the two only through named stepsTool limits are not published and can change

A reset experiment

Run step 1 at the start of two turns, with an idle gap between them, and write down both times. In the first turn, also write a small marker file to the VM's own disk, for example date > /tmp/marker.txt.

  • If the marker is still there in the second turn, the same VM was restored. The sandbox and its state quotes the documented pause and restore.
  • If the marker is gone, the VM was reclaimed and a fresh one started.
  • Do not rely on uptime alone. A restored VM may not count the time it spent paused, so a short uptime does not prove a new machine.

On November 8 and 9, the uptime was short and the demo file came back only as a copy from the conversation. The marker test was not run then. How long a paused VM is kept is still unknown.

Caveats

Claims in this section carry the labels Observed, Reported by ChatGPT, Tool-documented, Vendor-documented, Inferred or Unknown. The overview explains each one.

The observations come from one cloud Claude chat session on November 8 and 9, 2026, and from one conversation linked to a Mac on November 16, 2026. They may not hold for every plan, account or surface. Beyond Anthropic's published overview, the internal architecture cannot be seen from inside a session. Vendor behavior, feature flags and availability may change.

Sources

The full details of each source are in this page's source list.

  • Claude Desktop, Code tab docs (Anthropic): the environments Local, Cloud, SSH and WSL.
  • Panaversity's inspection of a cloud Claude session: a Python test run on November 8, 2026, and a process and environment inspection on November 9, 2026. The method is on this page.
  • Panaversity's inspection of both VMs in a Claude conversation linked to a Mac, November 16, 2026. The method is on this page.
  • ChatGPT's inspection of its own Python and terminal environment, November 9, 2026, as reported to Panaversity.
  • Agents API architecture (OpenAI API docs): the hosted harness, the environment and the application server as separate parts.
  • ChatGPT Work cloud security (ChatGPT docs): Work Cloud's own execution environment.
  • Claude Managed Agents overview (Claude Platform docs): the hosted harness and sandbox, sessions, events and beta status.
  • Managed Better Auth is now an OAuth provider, with MCP support (Neon, November 2026).
  • Better Auth MCP plugin (Better Auth docs).
  • Schedule recurring tasks in Claude Code Desktop (Claude Code docs): local, cloud and /loop tasks, missed runs, permissions and requiresUserInteraction.
  • Let Claude coordinate ongoing work with Projects (Claude Code docs): the coordinator, threads, standing context, sandbox resume, local threads, limits and beta status.
  • Projects redesigned: from folder to conversation (Anthropic blog, September 17, 2026).
  • Use Claude Code in the cloud (Claude Code docs): idle sessions, reclaimed VMs and what is restored.
  • Configure cloud environments (Claude Code docs): idle VMs pausing and being restored, network access levels, setup scripts and the environment cache, environment variables, network secrets, the GitHub proxy and shared environments.
  • Self-hosted environments (Claude Code docs): environments, runners and sessions, leases, what stays on your machines, model routing, availability and limits.
  • Verify session identity in self-hosted environments (Claude Code docs): the session token, its claims, what it proves and how to check it.
  • Message your other Claude Code sessions (Claude Code docs): how messages travel, what a message can and cannot do, inbound controls and limits.
  • How agent identity works, and Security and data handling (Claude Tag docs): service accounts, the credential store, access scope, sandboxes for each thread and availability.
  • Automate work with scheduled tasks (Claude Code docs): cloud runs as full Claude Code cloud sessions, fresh copies of repositories, auto mode and limits.
  • Claude Cowork architecture overview (Anthropic Help Center): the sandbox for each session, the egress proxy, credentials and plan availability.
  • Unlocking the Codex harness: how we built the App Server (OpenAI, February 2026).
  • Introducing ChatGPT agent (OpenAI, July 2025).
  • ChatGPT: What's new (OpenAI).
  • ChatGPT sandbox docs (OpenAI).
  • OpenAI is folding Codex into the ChatGPT app (The New Stack, July 2026).
  • Work with Claude Tag, overview (Claude Tag docs): Claude in team Slack channels, plans and availability.
  • Run agents in parallel (Claude Code docs): subagents, agent view, agent teams, dynamic workflows and projects compared.
  • Linux Foundation launches the Agent2Agent protocol project (Linux Foundation, June 23, 2025).
  • Tour of the Apps: the product features this section builds on.
  • Claude Desktop: Code tab docs (Anthropic, verified 2026-11-09)
  • Panaversity's inspection of a cloud Claude session: a Python test run on November 8, 2026, and a process and environment inspection on November 9, 2026
  • ChatGPT's inspection of its own Python and terminal environment, November 9, 2026, as reported to Panaversity
Owner
Panaversity