Agently Docs

Agently documentation for building AI applications with stable outputs, observable actions, and durable workflows.

View the Project on GitHub AgentEra/Agently

Agently 4.1.3.1 Release Notes

语言:English · 中文

Agently 4.1.3.1 是显式多轮任务信息管理的 foundation release。它加入了持久 Workspace 底座、Recall context skeleton,以及 Action Runtime 默认工作区继承能力, 让应用代码可以在多次执行步骤之间写入和取回任务信息。

这个版本不引入自主 WorkLoop planning。什么时候 observe、ingest、search、 checkpoint 和 build context,仍然由应用代码、TriggerFlow 定义或普通 Python loop 决定。

Highlights

agent = Agently.create_agent("issue-run").use_workspace("./.agently/runs/issue-123")

record = await agent.workspace.ingest(
    content={"route": None, "status": "failed"},
    collection="observations",
    kind="route_attempt",
    summary="Provider returned no route candidate",
    scope={"task_id": "issue-123", "area": "routing"},
    source={"type": "workflow", "step": "attempt-1"},
)

await agent.workspace.checkpoint(
    "issue-123",
    {"status": "failed", "evidence": record["id"]},
    step_id="attempt-1",
)

context = await agent.workspace.build_context(
    goal="Prepare the second routing attempt",
    scope={"task_id": "issue-123", "area": "routing"},
    budget={"max_items": 4},
)

当模型或 Action 层需要读写 Workspace 文件作业区时,单独暴露 file actions:

agent.use_workspace("./.agently/runs/issue-123")
agent.enable_workspace_file_actions(write=True)
agent.enable_shell(commands=["cat", "python"])

Action output 不会自动成为 memory,需要显式写入:

result = agent.action.execute_action("inspect_workspace_files", {"cmd": "cat notes/runtime.txt"})
await agent.workspace.ingest(
    content={"stdout": result["data"]["stdout"]},
    collection="observations",
    kind="action_output",
    summary="Shell inspection output",
    scope={"task_id": "issue-123"},
    source={"type": "action", "name": "inspect_workspace_files"},
)

Examples

Compatibility