Independent AI intelligence Two editions daily · ET
Fervor AI

Analysis · November 5, 2026 · concept

Conductoragent-harnessmulti-agentclaude-codecodex

Conductor's Shutdown Shows Which Half of Your AI Coding Workflow You Actually Own

The branches survive any workbench. The conversations, the setup and the reasons behind the code are another story.

Most developers running coding agents in parallel could tell you where their code lives. Far fewer could tell you where the reasoning behind that code lives, and this week that question stopped being academic.

Conductor, a popular Mac app for running several Claude Code and Codex agents side by side, announced it is joining Meta. New sign-ups stopped on November 4. Conductor Cloud goes read-only on November 13 and shuts down on November 20, when the team starts deleting user data. The local app keeps getting new models and critical fixes "until early 2027," then winds down too.

That is a short runway for a tool people built daily habits around. It also gives everyone else a free fire drill, because the same question applies to whatever workbench you use today: if it disappeared in two weeks, what would you still have?

The answer splits cleanly in two. Your code is safe almost by accident. The context around it is not, and context is the half that makes agent work worth keeping.

What actually happened, in plain terms

A coding-agent workbench is the app that sits between you and your agents. You give it a repo, it spins up separate copies of the project, starts an agent in each one, and shows you their progress and their diffs in one window. Conductor, Orca and a growing list of others all do some version of this.

Conductor's timeline, straight from its announcement:

  • Wednesday, November 4: no new sign-ups or downloads. Free users can't upgrade, and pro users keep pro features for free.
  • Friday, November 13: Conductor Cloud goes read-only. You can read old chats and export data, but you can't send new messages. Active subscribers get a refund.
  • Friday, November 20: Conductor Cloud shuts down and data deletion begins.
  • Early 2027: the local app starts winding down.

The team says it has "never sold or trained on Conductor user data," and that it "can't share much yet" about what comes next at Meta. Nobody needs to panic about the local app this month. But the cloud deadline is real, and November 13 is the date that matters, because that is when you lose the ability to keep working there.

The half you own: your code

Here is the good news, and it comes from a design choice that predates the AI era.

Conductor builds each workspace as a git worktree. A worktree is a second (or tenth) checkout of the same repository, each on its own branch, all sharing one history. Git's own docs put it simply: "A git repository can support multiple working trees, allowing you to check out more than one branch at a time" (git-worktree). Conductor's docs say its workspaces usually live under ~/conductor/workspaces/<repo name>/<workspace name> on your Mac.

Think of it like a shared kitchen with several cutting boards. Each agent gets its own board and its own ingredients, but there's one pantry. When the restaurant changes owners, the pantry stays in your house.

So for local workspaces, every commit and branch is already ordinary git. Run git worktree list in your main checkout to find the workspaces, and git status inside each one to spot unfinished changes. Those uncommitted edits are just files in that folder until you commit them, so commit or stash before you move anything. Cloud workspaces are different: those live on Conductor's servers until you pull them down, which is what the right-click "Transfer to local" option is for.

The half you rent: the context

Now the uncomfortable part. Look at what a day of agent work actually produces besides diffs:

  • The conversations. Every prompt you wrote, every plan the agent proposed, every "no, don't touch the migrations" correction. That transcript is the only record of why the code looks the way it does.
  • Setup knowledge. Which install command runs first, which port the dev server needs, which secret files get copied into a fresh workspace.
  • Personal settings. Your shortcuts, your preferred models, your custom actions.

Conductor splits these across places you control and places you don't. Shared project setup can live in a committed .conductor/settings.toml file, with commands like setup = "pnpm install" under a [scripts] table (Conductor docs). Per-machine overrides go in .conductor/settings.local.toml, and user settings sit in ~/.conductor/settings.toml. Workspace notes can go in a gitignored .context folder. And because git never copies ignored files into a new worktree, files like .env.local only show up if you list them under "Files to copy" or create them in the setup script.

That's a thoughtful design. It is also Conductor-shaped. A settings.toml written for one workbench is a recipe card in someone else's handwriting: readable, but no other app will follow it automatically.

The chats are the biggest exposure, because they live inside the app on both the cloud and local sides.

The clock you didn't know about

The workbench isn't the only thing deleting your history.

By default, Claude Code saves each session as a plaintext JSONL file (one JSON record per line) under ~/.claude/projects/, which is what makes claude --resume and claude --continue work (how Claude Code works). Those files don't stay forever. Claude Code deletes transcripts older than the cleanupPeriodDays setting, and "the default is 30 days" (the .claude directory). The same page notes that transcripts "are not encrypted at rest."

So even with no acquisition at all, a CLI conversation older than 30 days becomes eligible for deletion the next time the cleanup sweep runs. Codex keeps local session files too (its --ephemeral flag exists to skip "persisting session rollout files to disk," per the Codex CLI reference), and codex resume reopens them.

This is the real lesson of Conductor's notice. The workbench shutting down just made visible a gap that was already there: agent context has always been the disposable part of the workflow, and most of us never decided that on purpose.

Put this into practice

None of this takes long. Here is the order I'd go in, cheapest first.

1. If you use Conductor Cloud, move your work before November 13. Right-click each cloud workspace and choose Transfer to local. Then run the full export from Settings, Account, Export data. Do both: the transfer keeps you working, and the export is your archive if the local app goes away first.

2. Write project instructions into a file every agent reads. Claude Code reads CLAUDE.md at the start of every session, and recent versions will also read AGENTS.md when no CLAUDE.md exists (Claude Code memory docs). AGENTS.md is the open format Codex and many other agents use. If your conventions live in an app's settings panel, copy them into one of these files and commit it. Now any workbench, or a bare terminal, picks them up.

3. Turn setup into a plain script. Keep .conductor/settings.toml if you like it, but make it call something generic:

[scripts]
setup = "./scripts/setup.sh"

Then put the real steps (install, copy .env.example to .env.local, seed the database) in scripts/setup.sh. Orca, a shell alias or a new teammate can all run the same script. Expect the first version to miss a step; run it in a fresh worktree to find out which.

4. Make the agent write down why, as well as what. Before you merge, ask the agent to summarize its key decisions and rejected approaches into the pull request description or a docs/decisions/ note. Two paragraphs is plenty. That summary is the part of the transcript you'll want in six months, and it lives in git forever.

5. Keep Claude Code transcripts longer if you want them. Set a bigger retention window in your user settings (~/.claude/settings.json):

{ "cleanupPeriodDays": 180 }

The setting is listed in the settings reference. Remember these files are plaintext and unencrypted, so a longer window means more pasted secrets sitting on disk. Pick a number on purpose.

6. Learn the bare commands once. A workbench is a convenience over a few lines:

git worktree add ../myapp-auth-fix -b auth-fix
cd ../myapp-auth-fix && claude
git add -A && git commit -m "auth fix"   # commit before cleanup
git worktree remove ../myapp-auth-fix

git worktree remove refuses a folder with uncommitted changes, which is a feature: it stops you from deleting work by accident. You won't want to live like this. But knowing it works means no app can hold your workflow hostage.

7. Pick a fallback before you need one. Orca is an MIT-licensed desktop app that runs CLI coding agents side by side in separate worktrees on your own subscriptions, and it shipped v1.4.250 on November 5. Worktree-focused CLIs like worktrunk do the same job from the terminal. Try one on a throwaway repo this week so the switch isn't your first time.

Honest limitations

I haven't seen the contents of Conductor's export archive, so I can't tell you whether chats come out in a format another tool can read. Treat it as an archive, not a migration.

App-specific settings won't carry over anywhere. Even a perfectly committed .conductor/settings.toml is just a note to yourself once Conductor is gone.

Moving context into the repo has costs. Instruction files go stale, decision notes need a habit to stay useful, and a long AGENTS.md can make every agent session slower and noisier. Start small.

Transcripts are messy and sometimes sensitive. Keeping them longer means keeping whatever got pasted into them. For many teams, the right move is a short decision note in git and a modest retention window, not hoarding every session.

And this isn't a case against workbenches. Conductor was useful, Orca is useful, and so is the next one. The point is to use them the way you'd use a good rental: enjoy it, and keep your important things in boxes you can carry.

Where this leaves you

Conductor's notice is a sixteen-day warning for its users and a free audit for everyone else. Open your current setup and sort it into two piles: what lives in git, and what lives only in an app. The first pile survives anything. The second pile is a bet that a company's plans won't change.

You get to decide how big that bet is. Most of it can move into the repo this afternoon.

Sources: Conductor announcement; Conductor docs on git worktrees, workspaces and branches and project configuration; git-worktree documentation; Claude Code docs on how it works, the .claude directory, memory and settings; Codex CLI reference; AGENTS.md; stablyai/orca (v1.4.250, November 5, 2026); max-sixty/worktrunk. All read November 5, 2026.