Skip to main content

Buyer guide

Zed AI Rollout Playbook

How a team picks its agents and models in Zed, sets tool permissions, and reviews agent edits before merge.

Reading time
2 min read
Sources
18 cited
In this piece
Zed+25 Positive

Zed's agent panel runs the Zed Agent on Zed-hosted models, provider keys, existing subscriptions, gateways or local models, or an external agent from the ACP Registry. Without a shared policy, each engineer runs a different agent and model, and every external agent signs in, bills and handles data on its own terms. Agree which agents and model paths are allowed, set tool permission rules before agents run commands, and hold agent edits to the same review as any change. The steps follow Zed's documentation and what public accounts report.

When you need this

  • Engineers have moved to Zed and each one runs a different agent and model.
  • Claude Code or Codex users want to run their agent inside the editor.
  • Security asks what data leaves the machine and which commands an agent may run.
  • You are deciding whether Zed for Business is needed.

Evaluation criteria

  1. Before work: choose agents

    Which agents may engineers run in the panel?

    Look forThe Zed Agent plus a short list of external agents from the ACP Registry. Each external agent owns its own login, billing, and data terms, so review each one. Run the same task in the panel and in the agent's own app first. Public accounts report Claude Code hooks, sub-agent progress, and context controls missing in the panel.

  2. Before work: choose model paths

    Where does model traffic go?

    Look forZed-hosted models, provider keys kept in the system keychain, existing ChatGPT, Claude, or Copilot subscriptions, gateways such as OpenRouter or Bedrock, or local models. A corporate buyer on Hacker News uses own keys to keep traffic on a company OpenAI tenant with its budget controls.

  3. During work: set tool rules

    Which tool calls run without a prompt?

    Look foragent.tool_permissions with confirm as the default, always_allow patterns for safe build and test commands, and always_deny for destructive ones; deny patterns win even when the default is allow. MCP tools take the mcp:<server>:<tool> form.

  4. During work: run threads in worktrees

    Can parallel agents step on each other?

    Look forOne thread per task, each in its own linked Git worktree from the Threads Sidebar. Switch to the classic layout if engineers lose the file tree in the agentic one.

  5. Before merge: review every edit

    Does an agent edit get the same review as a colleague's?

    Look forAccept or reject hunks in the multibuffer, then normal pull request review. Practitioners who report steady results plan in ask mode, implement in write mode, run tests and diagnostics, and review line by line before opening a pull request.

  6. Before merge: set Business controls

    Do admins need to lock settings?

    Look forZed for Business at US$30 per seat per month to restrict hosted models, turn off edit predictions, and lock data sharing. These controls govern Zed's hosted services; SSO, SAML, and SCIM are planned.

Questions to ask vendors

  1. Which Claude Code and Codex commands does Zed's panel support in the release we deploy?
  2. When will SSO, SAML, and SCIM reach Zed for Business?
  3. How does a thread recover when an external agent hits a usage limit?
  4. Can external agents run in shared collaboration projects?

Red flags

  • Every engineer runs a different agent and model with no list of what is allowed.
  • Tool permissions left at allow with no deny patterns on machines holding production credentials.
  • A team moves off Cursor for Zed without piloting edit predictions on its own code.
  • Agent edits merge without a hunk review and a pull request review.

Products in this piece

Sources

Official documentation
10
Public source
8
  1. Zed docs: Agent PanelOfficial documentationzed.dev
  2. Zed docs: External AgentsOfficial documentationzed.dev
  3. Zed docs: Terminal ThreadsOfficial documentationzed.dev
  4. Zed docs: Parallel AgentsOfficial documentationzed.dev
  5. Zed docs: LLM ProvidersOfficial documentationzed.dev
  6. Zed docs: Tool PermissionsOfficial documentationzed.dev
  7. Zed docs: Model Context ProtocolOfficial documentationzed.dev
  8. Zed docs: AI PrivacyOfficial documentationzed.dev
  9. Zed for BusinessOfficial documentationzed.dev
  10. Zed pricingOfficial documentationzed.dev
  11. Hacker News, sodacanner: what Claude Code lacks inside Zed (Apr 29, 2026)Public sourcenews.ycombinator.com
  12. Hacker News, philipp-gayret: Claude Code hooks in Zed (Apr 9, 2026)Public sourcenews.ycombinator.com
  13. Alek Dobrohotov: Zed + Claude Code, a beautiful promise not quite there yet (Sep 6, 2025)Public sourcealekdobrohotov.substack.com
  14. Hacker News, sethhochberg: own keys for a corporate buyer (Sep 24, 2025)Public sourcenews.ycombinator.com
  15. Hacker News, trenchpilgrim: ask mode, write mode, then review (Aug 5, 2025)Public sourcenews.ycombinator.com
  16. Hacker News, senko: planning and building in one agent panel thread (Jul 21, 2025)Public sourcenews.ycombinator.com
  17. GitHub issue #55501: Claude Code ACP session stuck after a usage limit (May 2, 2026)Public sourcegithub.com
  18. GitHub issue #61343: external agents missing in shared projects (Jul 20, 2026)Public sourcegithub.com

Keep reading

  1. Buyer guideZed Models and Billing: Hosted Plan, Own Keys, or Agent SubscriptionsWhat each path to a model costs in Zed, which features each one keeps, and how to keep spend visible.
  2. ResearchZed's AI Features in Practice: What 86 Public Accounts ReportA coded read of public developer accounts of the Zed Agent, edit predictions, ACP agents, and AI pricing from May 2025 to October 2026.
  3. ReviewZed Review: A Native Editor That Hosts Your AgentsWho Zed fits as an AI editor, where it trails Cursor, and which agent, model, and billing choices to settle before a team moves.
Read the research