Temporal Agent Harness vs Orchestrator Zero

Yes

Agents run on Temporal workers, which run on hosts outside the Temporal service that you choose.

docs.temporal.io

Yes

Workers are Python processes on the Temporal Python SDK, which Temporal publishes as native wheels for Windows, macOS and Linux, and the harness itself is pure Python; its own docs say nothing about operating systems.

pypi.org

Yes

The Temporal service is open source and can be self-hosted.

docs.temporal.io

Yes

Workers poll task queues, so the service never calls into them.

docs.temporal.io

Conditional: Worker list, no machine join

Temporal lists and versions workers and routes tasks through queues. Managing host enrollment, revocation and operating-system upgrades remains outside the harness.

docs.temporal.io docs.temporal.io

Yes

Agents survive crashes and resume mid-turn where they left off.

github.com

Yes: Per agent and session

Each tool call is checked against a serialisable approval policy (allow lists, inherently safe tools, and an auto mode with approve, deny and escalate criteria), and anything not allowed waits for a person; the policy is set per agent, overridable per session and at runtime, and enforced in the workflow.

github.com temporal.io

Yes: Slack, Discord and Teams

A gated tool call pauses inside the workflow until a decision arrives from the web UI, the API, or the optional Slack and Discord front end in the repository; a Nexus connector covers Slack and Microsoft Teams.

github.com github.com github.com

Not documented

The MCP gateway discovers tools from registered servers. The docs do not describe a stored definition comparison that blocks changed tools pending approval.

github.com

Conditional: Namespaces, access is yours

Temporal namespaces separate work, but on a self-hosted service restricting access to each one needs a custom authorizer you write.

docs.temporal.io

Conditional: Tokens counted, no budget

The harness records token usage. A built-in monetary budget that stops a run is not documented.

github.com

No: Keys in worker environment

Model calls run as activities on your workers, which read the provider key from their environment; Temporal offers no gateway that holds it.

github.com

Conditional: Code Mode only, no allowlist

Model-written Code Mode scripts run in the Monty sandbox and act only through your tools, but the tools themselves run on the worker with no isolation or egress allowlist.

github.com

Conditional: History, not tamper-evident

Temporal history and the harness event stream record turns, tools and approval decisions. Cloud audit logs cover control-plane operations; a verifiable hash chain or signature for agent activity is not documented.

github.com docs.temporal.io

Conditional: Delete runs, no per-user keys

Each agent session is a workflow you can delete with its history, and payload encryption is a codec you write and key yourself, so per-person keys and crypto-shredding are your own code.

docs.temporal.io docs.temporal.io

Conditional: Retention, no legal hold

A namespace sets how long closed workflow histories are kept, 1 to 90 days on Temporal Cloud, and nothing documents a legal hold.

docs.temporal.io docs.temporal.io

Yes

Integrations cover Google Gemini, the OpenAI Agents SDK (OpenAI and LiteLLM models) and Pydantic AI, whose providers include Anthropic and local models; the harness docs do not name local models themselves.

temporal.io github.com github.com

Conditional: MCP, no A2A

Agents use MCP servers through the OpenAI Agents integration and a Nexus MCP package; agents call each other through the harness's own typed interfaces, and A2A does not appear in the repository.

github.com github.com