Temporal Alternative Without the Cluster and Determinism Constraints

#temporal #python #durableexecution #workflow

Temporal is the gold standard for durable execution. If you need long-running workflows that survive crashes, it's the first thing most teams evaluate.

But then you read the docs. And you discover what Temporal actually requires.

The Cluster Problem

Temporal needs a cluster. Either you run it yourself (Temporal Server + Cassandra/PostgreSQL + Elasticsearch) or you pay for Temporal Cloud.

Self-hosted means:

Temporal Server (3+ nodes for HA)
Cassandra or PostgreSQL for persistence
Elasticsearch for visibility
Monitoring, upgrades, schema migrations
A team that understands Temporal internals when something breaks at 2am

This is fine if you're Uber. If you're a team of 5 building an AI agent pipeline, it's a lot of infrastructure for "I want my workflow to survive a crash."

The Determinism Problem

Temporal replays your workflow code on every restart. This means your workflow functions must be deterministic. No side effects.

# These all break Temporal workflows:
import random
random.randint(1, 100)       # non-deterministic

from datetime import datetime
datetime.now()                # different on replay

import requests
requests.get("https://api.example.com")  # side effect

Every developer on the team needs to learn this. New hire writes datetime.now() in a workflow, the replay breaks in production, and nobody understands why until someone reads the Temporal determinism docs.

Activities solve this - you put non-deterministic code in activities. But that means restructuring your code around Temporal's execution model. Your agent code now has to know it's running inside Temporal.

What If You Just Didn't

from axme import AxmeClient, AxmeClientConfig

client = AxmeClient(AxmeClientConfig(api_key=os.environ["AXME_API_KEY"]))

intent_id = client.send_intent({
    "intent_type": "intent.pipeline.process.v1",
    "to_agent": "agent://myorg/production/data-pipeline",
    "payload": {
        "steps": ["extract", "validate", "transform", "load"],
        "source": "postgres-main",
        "destination": "warehouse",
    },
})
result = client.wait_for(intent_id)

No cluster. No determinism constraints. Write normal Python. Call datetime.now() all you want.

The state lives in the platform (managed PostgreSQL). Your agent is stateless. If it crashes, the platform redelivers the intent. If it needs human approval mid-workflow, the platform handles the wait.

Side-by-Side

	Temporal	AXME
Infrastructure	Cluster (self-hosted or Cloud)	Managed API
Determinism constraints	Required for workflow code	None
Learning curve	Weeks (activities, signals, queries, replay)	Hours
Human approval	Build it (signals + UI + notifications)	Built-in
Crash recovery	Replay-based (determinism required)	Redelivery-based (stateless agent)
Setup time	Days-weeks	`pip install axme`