flatmachines 2.2.0

State machine orchestration for agent workflows.

FlatMachines (Python SDK)

State-machine orchestration for LLM agents. FlatMachines is agent-framework agnostic and uses adapters to execute agents from FlatAgents, smolagents, pi-mono, or other runtimes.

For LLM/machine readers: see MACHINES.md.

Install

pip install flatmachines[flatagents]

Optional extras:

• flatmachines[cel] – CEL expression engine
• flatmachines[validation] – JSON schema validation
• flatmachines[metrics] – OpenTelemetry metrics
• flatmachines[smolagents] – smolagents adapter

Quick Start

from flatmachines import FlatMachine

machine = FlatMachine(config_file="workflow.yml")
result = await machine.execute(input={"query": "..."})
print(result)

workflow.yml

spec: flatmachine
spec_version: "0.10.0"

data:
  name: reviewer
  agents:
    reviewer: ./reviewer.yml
  states:
    start:
      type: initial
      agent: reviewer
      input:
        code: "{{ input.code }}"
      output_to_context:
        review: "{{ output }}"
      transitions:
        - to: done
    done:
      type: final
      output:
        review: "{{ context.review }}"

Agent Adapters

FlatMachines delegates agent execution to adapters. Built-ins are registered automatically if their dependencies are installed:

• flatagent – uses FlatAgents configs (flatmachines[flatagents])
• smolagents – executes MultiStepAgent via agent_ref.ref factories
• pi-agent – Node bridge to pi-mono using pi_agent_runner.mjs

Agent refs in data.agents can be:

• string path to a flatagent config
• inline flatagent config (spec: flatagent)
• typed adapter ref: { type: "smolagents" | "pi-agent" | "flatagent", ref?: "...", config?: {...} }

Custom adapters can be registered via AgentAdapterRegistry.

State Execution Order

For each state, the Python runtime executes in this order:

1. action → hooks.on_action
2. launch → fire-and-forget machine(s)
3. machine / foreach → peer machines (blocking)
4. agent → adapter + execution strategy
5. output → render final output for type: final

Jinja2 templates render input, context, and output. A plain string like context.foo resolves to the actual value (not a string).

Execution Types

execution:
  type: retry
  backoffs: [2, 8, 16]
  jitter: 0.1

Supported: default, retry, parallel, mdap_voting.

Parallelism & Launching

• machine: child → invoke peer machine (blocking)
• machine: [a, b] → parallel
• foreach: "{{ context.items }}" → dynamic parallelism
• launch: child → fire-and-forget; result is written to backend

Persistence & Resume

persistence:
  enabled: true
  backend: local  # local | memory
  checkpoint_on: [machine_start, state_enter, execute, state_exit, machine_end]

Resume with machine.execute(resume_from=execution_id). Checkpoints store MachineSnapshot including pending launches.

Hooks

Configure hooks via data.hooks:

hooks:
  file: ./hooks.py
  class: MyHooks
  args: { ... }

Built-ins: LoggingHooks, MetricsHooks, WebhookHooks, CompositeHooks.

Invokers & Result Backends

Invokers define how peer machines are launched:

• InlineInvoker (default)
• QueueInvoker (base class for external queues)
• SubprocessInvoker (python -m flatmachines.run)

Results are written/read via ResultBackend URIs: flatagents://{execution_id}/result.

Logging & Metrics

from flatmachines import setup_logging, get_logger
setup_logging(level="INFO")
logger = get_logger(__name__)

Env vars match FlatAgents: FLATAGENTS_LOG_LEVEL, FLATAGENTS_LOG_FORMAT, FLATAGENTS_LOG_DIR.

Metrics require flatmachines[metrics] and FLATAGENTS_METRICS_ENABLED=true.

CLI Runner

python -m flatmachines.run --config machine.yml --input '{"key": "value"}'

Examples (Repo)

• character_card
• coding_agent_cli
• custom_coding_workflow
• dfss_deepsleep
• distributed_worker
• dynamic_agent
• error_handling
• gepa_self_optimizer
• helloworld
• human-in-the-loop
• mdap
• multi_paper_synthesizer
• parallelism
• peering
• research_paper_analysis
• rlm
• story_writer
• support_triage_json
• writer_critic

Specs

Source of truth:

• flatmachine.d.ts
• flatagent.d.ts
• profiles.d.ts