AXMEAXME

AXME CLOUD

Write agent workflows without determinism constraints — AXME handles replay differently

Temporal's determinism requirement breaks the moment you introduce an LLM call — you can't call OpenAI inside a Workflow function. This is the #1 adoption blocker for AI teams.

Unlike Temporal, AXME's intent lifecycle model doesn't require deterministic code.

Agents are not replayable microservices

Temporal and similar engines assume workflow code is deterministic so the platform can replay history. The moment you call an LLM, use randomness, or depend on live external APIs, that model breaks — teams split logic into Activities, sagas, and workarounds.

AXME tracks state at the intent lifecycle boundary. Non-deterministic code runs inside your agent as usual; durability comes from persisted intent state, not from re-executing every line of Python.

CAPABILITIES

How it works.

Intent-level state

Replay at lifecycle boundaries, not every line.

LLM-friendly

Non-deterministic code is first-class.

External I/O

APIs, randomness, humans — no workaround layers.

LLM calls inside workflows

Temporal

# Cannot call OpenAI inside Workflow fn
# Must use Activities + strict replay rules

AXME

intent = axme.submit(agent_step)
result = await openai.chat(...)  # OK
await intent.complete(result)

When Temporal still wins

Mature microservice orchestration with existing Temporal expertise and purely deterministic workflows.

When to choose AXME vs Temporal

Choose AXME when agent steps include LLMs, human gates, and heterogeneous tools in one flow. Temporal remains strong for mature deterministic microservice orchestration where teams already invested in worker fleets and replay discipline.

Migration paths often run AXME for new agent features while legacy Temporal workflows wind down — see /migrate/from-temporal/ for concept mapping.

Common questions

Can I still use idempotent tool calls?
Yes — idempotency at the tool layer is good practice. AXME does not require deterministic functions; it requires durable intent state.
How are failures recovered?
Failed intents retain context for retry, human escalation, or inspection — without replaying the entire workflow from line one.
Is randomness in agent code allowed?
Yes — random seeds, sampling, and LLM temperature do not violate the execution model.

Related

Related links

Ship your first durable agent — in under 10 minutes.

Free tier. No credit card. Self-host or hosted — your choice.

Start free now Read the docs