Agentic software should feel like software.
AgentDeck is the execution layer for agentic software. You write agents, tools and workflows as
ordinary Python; AgentDeck gives them one execution model underneath, with sessions, streaming,
a canonical event log, human approval and run control already in place. One run is one run whether
it started from the terminal, over HTTP or through AG-UI, and it can be observed, paused, answered
or cancelled from anywhere.
pip install agentdeck-sdk
from agentdeck import Agent, Deck
deck = Deck(agents=[Agent(name="assistant", model="gpt-4o-mini")])
async with deck:
run = await deck.runs.start("assistant", input="Hello!")
async for event in run.events(follow=True):
print(event.kind)Python 3.12+, MIT licensed.
|
|
| Build |
agents, tools and workflows as plain Python. Deck is the one composition root, and it validates dependencies, schemas and catalogs before the first turn. |
| Run |
one execution model for every target, with sessions that survive turns, surfaces and restarts. |
| Observe |
one canonical event log. Traces, cost and audit are readers of it, not per-engine instrumentation. |
| Control |
pause, resume, cancel or answer any run by id, from any process, durably across restarts. ctx.ask() parks a run on a human. |
| Connect |
serve the same Deck over HTTP/SSE, the terminal and AG-UI at once. Bring MCP servers and agents you already have. |
| Repository |
|
| agentdeck |
the SDK: source, issues, releases and the full documentation set |
Docs · PyPI · Releases · Changelog
