auctorai/durable_agents

Durable Agents is a runnable demo project that accompanies a blog post exploring how AI agents can be built as self writing workflows. It shows how to connect Pydantic AI, a Python framework for building AI powered applications, with Temporal, a workflow orchestration system that makes long running processes resilient and restartable, so that every step an AI agent takes becomes a durable, retryable unit of work. The core idea is that when an AI model decides to use a tool, that tool call is automatically converted into a Temporal activity, which is a tracked unit of work that Temporal can retry on failure and resume after a crash. The project also supports spawning subagents, where the model starts a child workflow for a subtask, and programmatic tool calling, where the model writes a short script that orchestrates multiple tools in parallel or in sequence. The demo runs without any AI API key by default. A scripted model replaces a real language model and emits pre defined tool calls, making it easy to explore workflow behavior deterministically. You can switch to a live model by providing an OpenAI or Anthropic API key instead. A built in REPL lets you type prompts and watch the resulting Temporal workflow graph, and it leaves any file side effects in a local sandbox folder for inspection. Error handling is deliberate. Validation errors are fed back to the model as correctable feedback, known execution failures come back as terminal errors, retryable infrastructure failures are retried according to Temporal's policy, and parent cancellation stops all child work in progress without triggering another model turn. The test suite covers all these scenarios using Temporal's in-process test environment, checking that tool calls, subagent spawning, and programmatic tool calling all produce the expected workflow shape.