The harness
The harness is the agent runtime that runs inside the sandbox. Claude Code, Codex, and OpenCode are supported. Cursor is supported when enabled for the account. Hermes and OpenClaw appear as coming soon. Pick the harness per run in Agent Studio, save it on an agent config, or choose it for Signal under Configure Signal. Every harness needs a model credential, which Arize calls an AI provider:
Codex, Cursor, and OpenCode run only in the Arize sandbox. Claude Code can also run in a Claude Managed Agent environment when that option is enabled. Vercel Sandbox and Daytona are coming soon.
Enterprise accounts can bring their own key, and Cursor requires one.
1
Add an AI Provider
Under Settings → Account Settings → AI Providers, add a provider supported by the selected harness.
2
Allow it in agents
Tick the Managed Agents checkbox on the provider’s row in the AI Providers table. You can also do it from the AI Provider picker itself: providers appear under Enabled for managed agents or Not enabled for managed agents, and the latter have an inline Enable Key action.
3
Select it
Choose that provider on an agent config, during agent setup, when enabling Signal, or as the space default under Space Settings.
The harness picker appears in Agent Studio, on agent configs, and (on Enterprise) under Configure Signal. Signal defaults to Claude Code, and the AI provider you choose for Signal has to match the harness it runs on.
Skills
Skills are account-level integrations that give sandbox workers access to external systems. They are shown under Space Settings, and you connect them at the Connect step of agent setup. A saved skill can be attached to multiple managed agent configs and works with every harness.Arize Skills are a different feature that runs in your own coding agent on your own machine. Managed agent skills inject credentials into an Arize-managed sandbox.
GitHub
Lets a worker clone a repo, branch, and open PRs. Secret:GH_TOKEN. Grant only the repository permissions required for the task. Opening pull requests requires repository write access, and modifying workflow files may require additional workflow permissions. Arize verifies that the token is present but does not validate its GitHub scopes.
Classic personal access token scopes:
Create the token at github.com/settings/tokens with Generate new token (classic), then paste it into a GitHub skill under Settings → Space Settings → Skills, or from Configure Signal → Attach Repo when wiring Signal to a repo (
owner/repo).
If your account offers Install GitHub App, use that instead of a PAT.
Arize
This skill is optional, because every run already receives a temporary, space-scoped Arize credential governed by its agent role. Attach an Arize skill only when you want the run to use a supplied API key instead. Secret:ARIZE_API_KEY. The worker can only call what that key allows, so use a dedicated key with the minimum permissions the task needs. Do not use a personal admin key.
Dynatrace
Lets a worker query Dynatrace alongside Arize traces. Secrets:DT_API_TOKEN and DT_ENVIRONMENT_URL. The URL must be an HTTPS tenant root such as https://abc123.apps.dynatrace.com or https://abc123.live.dynatrace.com, without a path, query, port, or embedded credentials.
Datadog
Lets a worker query Datadog metrics and logs alongside Arize traces. Secrets:DD_API_KEY and DD_APP_KEY. DD_SITE is optional.
Custom skill
Installs a skill from a GitHub repository into the sandbox, for example an internal runbook or a fork of an Arize skill.
Custom skills may declare optional env vars, while every non-custom provider requires at least one secret.
Add a skill
1
Start agent setup
Launch an agent from Signal & Agents → New Agent. The Connect step lists the skills that agent needs.
2
Choose a provider
GitHub, Arize, Dynatrace, Datadog, or Custom skill.
3
Enter secrets
Paste them as
ENV=value lines or a JSON object. The form shows which env vars it will store.4
Save
The skill becomes available to managed agent configs in the account. For GitHub, set the repo after saving.
What a worker can reach
The automatic Arize credential is limited to the selected space and its agent role. Attached skills can reach any resource their credentials authorize.
Each run gets a dedicated sandbox that is torn down when the run ends. Agent execution is traced, so you can audit tool use and outputs from the run’s detail view.
What a worker cannot do
A managed agent is not a customer-facing production runtime. It improves the systems you observe in Arize and does not replace your AI agent. It also does not guarantee that a person reviews its changes. For repository changes, instruct the agent to open a pull request and review it before merging, since other credentials may permit direct mutations. Its automatic Arize credential cannot exceed the permissions of the selected agent role, and attached skills remain governed by their own credentials.Permissions
Account Admins grant managed-agent permissions on custom roles at the space level. Under legacy space roles, Space Member+ has full managed-agent access, and Read-only can view Signal issues but not agent jobs or transcripts.
Each config can select an agent role, which config-backed runs inherit, and ad hoc runs can select one directly. If no role is selected, Arize uses Managed Agent (Default).
Security practices
Keep each credential as narrow as the task allows. Give GitHub tokens read-only access until you need pull requests, and scope Dynatrace tokens to the APIs in use. Create one skill per integration and reuse it across agent configs so the same secret is never stored twice. Credentials belong on the skill or the AI Provider integration, never in prompts. Skill secrets are encrypted at rest and are not returned in plaintext to the browser, and they are made available to the selected integration when a run executes. Install custom skills only from sources you trust, because the sandbox clones and executes custom skill code during a run. Treat the PRs and investigation suggestions a run produces as untrusted automation output until a person has read them.Related docs
- Connect your AI provider: the Arize-managed model, your own key, or your own gateway, and where your data goes
- Managed agents: the gallery, agent configs, and customization
- Cost agent: the ready-made agent for spend
- Signal: ranked issues per tracing project
- Custom Roles: grant
MANAGED_AGENT_*permissions - Managed Agents CLI: skills: create, rotate, and delete skills from the terminal with
ag skill - Anthropic AI Provider: bring your own Anthropic key for Claude Code
- OpenAI AI Provider: bring your own OpenAI key for Codex
- Cursor AI Provider: add the Cursor key that powers the Cursor harness