Temporal introduces an early, open-source Agent Harness that wraps around existing agent SDKs (OpenAI Agents SDK, PydanticAI, Gemini) to add durable execution, tool-call approval policies, typed agent interfaces, multi-turn orchestration, a Code Mode for running model-generated Python durably, callback tools for out-of-process actions, and a structured AgentEvent stream for observability, replay, and event-driven integration. The project runs on Temporal Workflows, giving agents crash-resilient, resumable execution and a durable history of everything they've done. It is available now on GitHub in an early, pre-preview state, with the team inviting feedback on abstractions and requests for additional inner-harness integrations.
Table of contents
But aren't there already enough harnesses out there? #Durable execution #Bring your own inner harness #Controls around tool calls #Turns, not just agent loops #Strongly typed agents #Agents compose like software #Durable, secure, and observable code mode #Tools can run where they need to #Beyond just features, an agent-native foundation #What we're sharing today #Come build with us #Questions this post answers
What is the Temporal Agent Harness and which agent SDKs does it support?
It is an early-stage, open-source outer harness from Temporal that wraps around existing agent SDKs to add durable execution, tool-call approval policies, typed interfaces, and multi-agent composition. It currently integrates with three inner harnesses: Google Gemini, the OpenAI Agents SDK, and PydanticAI, with more planned. It runs every agent as a Temporal Workflow. Engineers evaluating durable agent infrastructure follow releases like this one on daily.dev.
How does the Temporal Agent Harness handle human approval for risky tool calls like refunds?
It provides a policy layer around tool execution that can require human approval before specific tools run, pause the agent durably while waiting, and resume execution whenever the decision arrives, even hours or days later. Policies can be layered, changed at runtime, and scoped to specific tools or sessions, such as allowing safe reads but requiring approval for refunds. daily.dev helps developers building agent approval workflows keep up with tools like this.
What is Code Mode in the Temporal Agent Harness?
Code Mode lets a model write a Python program over a provided set of tools instead of making tool calls one at a time, using normal control flow like loops, conditions, and concurrency to orchestrate many tool calls within a single turn. That Python program remains durable, observable, and subject to the same approval policies defined for the agent, rather than being treated as an unmonitored sandbox. daily.dev surfaces developments like this for teams designing safer agent tool execution.