semora

Durable agent runtime for Python. This is the core: contracts, control points, tool execution, the orchestrator, the plain async while planner, and AgentRuntime.

By default AgentRuntime drives the planner directly — no orchestrator, no ledger:

from semora import Agent, AgentRuntime, ChatModel

model = ChatModel(model="gpt-4.1")
agent = Agent("reviewer", "Reviews repositories", model, tools, system_prompt)
outcome = await AgentRuntime().run("attempt-42", agent, "inspect this repository")

This path cannot suspend or recover a crashed round. Attach a ledger when those guarantees are required:

from semora import AgentRuntime, MemorySteps

runtime = AgentRuntime(store=MemorySteps())
outcome = await runtime.run("attempt-42", agent, "inspect this repository")

Policy lives on Controls / ControlPlane (on_inputs, before_model, pre_tool_use, post_tool_use, before_finish, on_resume, on_suspend). before_finish can refuse an ending and send the loop around again.

semora_coding.builtins.builtin_tools() supplies read, write, edit, grep, glob, Bash, and web_fetch. web_search is not included. Bash stays disabled until ExecToolOptions.allow_list is set. File and process effects use the WorkspaceProvider injected by AgentRuntime (semora.workspace).

Install extras beside it:

See the repository README for examples.