Agents that keep running. Noeta is a Python SDK for agents that survive
crashes, wait days for a human, and scale from one script to a cluster —
without changing the agent.
English · 简体中文 · Docs · Quickstart · Why Noeta
A working agent in a few lines
uv pip install noeta-sdk
export ANTHROPIC_API_KEY=sk-ant-...
from noeta.sdk import Options, query
from noeta.sdk.providers import AnthropicProvider
result = query(
Options(system_prompt="You are a concise coding assistant."),
goal="What files are in this directory, and what does each one do?",
provider=AnthropicProvider(),
model="claude-sonnet-5",
)
print(result.answer())
The agent explores the directory with its built-in file tools and answers.
result also holds every model call, tool call and token count from the run.
Python 3.11+. The Glob and Grep tools need ripgrep
on PATH (apt install ripgrep, brew install ripgrep).
What sets it apart
Survives crashes. Task state is rebuilt from an append-only event log,
never held in memory. Kill a worker mid-task and another one carries on from
the last step. The same log is a complete audit trail.
Waiting costs nothing. A task can pause for a human approval, a timer, a
subtask or an outside event — for seconds or months. It uses no resources
while it sleeps and is woken exactly once.
From a script to a cluster. query() in a script, Client.start_workers(n)
inside a service, several hosts on one Postgres. The agent code is the same at
every step, and there is no daemon or extra service to run.
Also: every capability is a plugin — file tools, web, memory, MCP,
sandboxes and model adapters use the same public API yours does. Any
model — Anthropic, any OpenAI-compatible gateway, or the OpenAI Responses
API, one line to switch. Approval before action — risky tool calls pause
for a human; guards can block a call before it runs.
How it compares
| Noeta | Claude Agent SDK | LangGraph | Temporal |
|---|
| What it is | Durable agent runtime, as a library | Agent loop library for Claude | Graph-based agent framework | Durable workflow platform |
| Control flow | The model decides each step | The model decides | A graph you define | Workflow code you write |
| What is saved | Every event; state is derived from it | The conversation | Checkpoints of graph state | Workflow history |
| Waiting for a human / timer | Built in, woken exactly once | Resume the conversation | Interrupt, then the caller resumes | Built in |
| Scaling out | Worker pool; many hosts on Postgres | One process | Up to you, or the hosted platform | A Temporal cluster |
| Models | Any, one line to switch | Claude | Any | — |
| Extra service to run | None | None | None | Temporal server |
Pick Noeta when an agent runs unattended for a long time and you need to
recover, audit and scale it. Full comparison, and when not to use it.
Benchmarks
Measured on public benchmarks — one run each, over a sample:
| Benchmark | Scope | Measured on | noeta-agent main (Claude Opus 4.8) | Field (full set) |
|---|
| Terminal-Bench 2.1 | 40-task stratified sample | noeta-sdk 0.6.28, 2026-09-19 | first pass 24/40; 33/40 best of 3 attempts | public board spans 58.7%–83.8% |
| SWE-bench Verified | 15-instance subset | noeta-sdk 0.6.10, 2026-08-09 | 13/15 after re-running 4 setup timeouts (first pass 9/15) | not comparable to a subset |
An agent built only on the public SDK (noeta-agent),
run on the official harness (harbor)
and scored by each task's own verifier. Both rows are single runs over samples,
not full leaderboard runs, and the higher figures count re-runs of tasks the
first pass missed. Method and caveats.
Documentation
Prefer code? examples/ has runnable scripts for custom tools,
MCP servers, permission gates, subagents and surviving kill -9, each with an
offline test. examples/reference-host/ is a
complete host built from the public API alone.
License
Apache 2.0 — see LICENSE.
