Agents

An agent turn is expensive and it touches real systems. Tokens cost money, and a tool call charges a card or sends an email. Losing the process partway through should not mean paying for those twice.

pyrula.agents runs agents on the durable workflow engine, so a turn interrupted halfway resumes from where it stopped: completed LLM calls and tool calls are replayed from the journal instead of re-issued. You write agents and tools, stream their output, and serve them over HTTP; the durability comes from the engine underneath.

from pyrula.agents.decorators.agent import agent
from pyrula.agents.decorators.tool import tool

Start with the Quickstart to define an agent and prove a crash resumes without re-running work. For the execution model agents inherit (runs, replay, suspend points), read the Workflows concepts.

Where to go

• Authoring: the @agent and @tool decorators, calling an LLM, and the context surface.
• Integrations: bring LangChain, PydanticAI, OpenAI Agents SDK, and CrewAI tools and agents into a Pyrula turn.
• Serving: run agents over HTTP, talk to them with the client, scale execution on workers, drive it from the CLI, and fire them on schedules.
• Regression testing: run local-first agent evals against the durable production path and emit CI-friendly JSON artifacts.
• Observability: OpenTelemetry tracing and metrics with GenAI semantic conventions.