GitHub - memgrafter/flatagents: Flat, declarative agents and state machines for orchestrating LLMs.

FlatAgents

Define LLM agents in YAML. Run them anywhere.

For LLM/machine readers: see MACHINES.md for comprehensive reference.

Why?

Composition over inheritance — compose stateless agents and checkpointable machines
Compact structure — easy for LLMs to read and generate
Simple hook interfaces — escape hatches without complexity; webhook ready
Inspectable — every agent and machine is readable config
Language-agnostic — reduce code in any particular runtime
Common TypeScript interface — single schema for agents, single schema for machines
Limitations — machine topologies can get complex at scale

Inspired by Kubernetes manifests and character card specifications.

Versioning

All specs (flatagent.d.ts, flatmachine.d.ts, profiles.d.ts) and SDKs (Python, JS) use lockstep versioning. A single version number applies across the entire repository.

Core Concepts

Use machines to write flatagents and flatmachines, they are designed for LLMs.

Term	What it is
FlatAgent	A single LLM call: model + prompts + output schema
FlatMachine	A state machine that orchestrates multiple agents, actions, and state machines

Use FlatAgent alone for simple tasks. Use FlatMachine when you need multi-step workflows, branching, or error handling.

Examples

Example	What it demonstrates
helloworld	Minimal setup — single agent, single state machine
writer_critic	Multi-agent loop — writer drafts, critic reviews, iterates
story_writer	Multi-step creative workflow with chapter generation
human-in-the-loop	Pause execution for human approval via hooks
error_handling	Error recovery and retry patterns at state machine level
dynamic_agent	On-the-fly agent generation from runtime context
character_card	Loading agent config from character card format
mdap	MDAP voting execution — multi-sample consensus
gepa_self_optimizer	Self-optimizing prompts via reflection and critique
research_paper_analysis	Document analysis with structured extraction
multi_paper_synthesizer	Cross-document synthesis with dynamic machine launching
support_triage_json	JSON input/output with classification pipeline
parallelism	Parallel machines, dynamic foreach, fire-and-forget launches

Quick Start

pip install flatagents[all]

from flatagents import FlatAgent

agent = FlatAgent(config_file="reviewer.yml")
result = await agent.call(query="Review this code...")
print(result.output)

Example Agent

reviewer.yml

spec: flatagent
spec_version: "0.8.3"

data:
  name: code-reviewer

  model: "smart-expensive"  # Reference profile from profiles.yml

  system: |
    You are a senior code reviewer. Analyze code for bugs,
    style issues, and potential improvements.

  user: |
    Review this code:
    {{ input.code }}

  output:
    issues:
      type: list
      items:
        type: str
      description: "List of issues found"
    rating:
      type: str
      enum: ["good", "needs_work", "critical"]
      description: "Overall code quality"

What the fields mean:

spec/spec_version — Format identifier and version
data.name — Agent identifier
data.model — Profile name, inline config, or profile with overrides
data.system — System prompt (sets behavior)
data.user — User prompt template (uses Jinja2, {{ input.* }} for runtime values)
data.output — Structured output schema (the runtime extracts these fields)

Model Profiles

Centralize model configurations in profiles.yml and reference them by name:

profiles.yml

spec: flatprofiles
spec_version: "0.8.3"

data:
  model_profiles:
    fast-cheap:
      provider: cerebras
      name: zai-glm-4.6
      temperature: 0.6
      max_tokens: 2048

    smart-expensive:
      provider: anthropic
      name: claude-3-opus-20240229
      temperature: 0.3
      max_tokens: 4096

  default: fast-cheap      # Fallback when agent has no model
  # override: smart-expensive  # Uncomment to force all agents

Agent usage:

# String shorthand — profile lookup
model: "fast-cheap"

# Profile with overrides
model:
  profile: "fast-cheap"
  temperature: 0.9

# Inline config (no profile)
model:
  provider: openai
  name: gpt-4
  temperature: 0.3

Resolution order (low → high): default profile → named profile → inline overrides → override profile

Output Types

output:
  answer:      { type: str }
  count:       { type: int }
  score:       { type: float }
  valid:       { type: bool }
  raw:         { type: json }
  items:       { type: list, items: { type: str } }
  metadata:    { type: object, properties: { key: { type: str } } }

Use enum: [...] to constrain string values.

Multi-Agent Workflows

For orchestration, use FlatMachine (full docs in MACHINES.md):

from flatagents import FlatMachine

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

FlatMachine provides: state transitions, conditional branching, loops, retry with backoff, and error recovery—all in YAML.

Features

Checkpoint and restore
Python SDK (TypeScript SDK in progress)
MACHINES.md — LLM-optimized reference docs
Decider agents and machines
On-the-fly agent and machine definitions
Webhook hooks for remote state machine handling
Metrics and logging
Error recovery and exception handling at the state machine level
Parallel machine execution (machine: [a, b, c])
Dynamic parallelism with foreach
Fire-and-forget launches for background tasks

Planned

Distributed execution — cross-network machine peering, inter-machine strategies
SQL persistence backend
TypeScript SDK
max_depth config to limit machine launch nesting
Checkpoint pruning to prevent storage explosion
$root/ path prefix — resolve agent/machine refs from workspace root, not config dir
Input size validation — warn when prompt exceeds model context window
Serialization warnings — flag non-JSON-serializable context values before checkpoint

Specs

TypeScript definitions are the source of truth:

Python SDK

pip install flatagents[litellm]

LLM Backends

from flatagents import LiteLLMBackend, AISuiteBackend

# LiteLLM (default)
agent = FlatAgent(config_file="agent.yml")

# AISuite
backend = AISuiteBackend(model="openai:gpt-4o")
agent = FlatAgent(config_file="agent.yml", backend=backend)

Hooks

Extend machine behavior with Python hooks:

from flatagents import FlatMachine, MachineHooks

class CustomHooks(MachineHooks):
    def on_state_enter(self, state: str, context: dict) -> dict:
        context["entered_at"] = time.time()
        return context

    def on_action(self, action: str, context: dict) -> dict:
        if action == "fetch_data":
            context["data"] = fetch_from_api()
        return context

machine = FlatMachine(config_file="machine.yml", hooks=CustomHooks())

Available hooks: on_machine_start, on_machine_end, on_state_enter, on_state_exit, on_transition, on_error, on_action

Execution Types

execution:
  type: retry              # retry | parallel | mdap_voting
  backoffs: [2, 8, 16, 35] # Seconds between retries
  jitter: 0.1              # ±10% random variation

Type	Use Case
`default`	Single call
`retry`	Rate limit handling with backoff
`parallel`	Multiple samples (`n_samples`)
`mdap_voting`	Consensus voting (`k_margin`, `max_candidates`)

Schema Validation

from flatagents import validate_flatagent_config, validate_flatmachine_config

warnings = validate_flatagent_config(config)
warnings = validate_flatmachine_config(config)

Logging & Metrics

from flatagents import setup_logging, get_logger

setup_logging(level="INFO")  # Respects FLATAGENTS_LOG_LEVEL env var
logger = get_logger(__name__)

Env vars: FLATAGENTS_LOG_LEVEL (DEBUG/INFO/WARNING/ERROR), FLATAGENTS_LOG_FORMAT (standard/json/simple)

For OpenTelemetry metrics:

pip install flatagents[metrics]
export FLATAGENTS_METRICS_ENABLED=true

Metrics are enabled by default and print to stdout every 5s. Redirect to file or use OTLP for production:

# Metrics print to stdout by default
python your_script.py

# Save to file
python your_script.py >> metrics.log 2>&1

# Disable if needed
FLATAGENTS_METRICS_ENABLED=false python your_script.py

# Send to OTLP collector for production
OTEL_METRICS_EXPORTER=otlp \
OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317 \
python your_script.py

Env vars for metrics:

Variable	Default	Purpose
`FLATAGENTS_METRICS_ENABLED`	`true`	Enable OpenTelemetry metrics
`OTEL_METRICS_EXPORTER`	`console`	`console` (stdout) or `otlp` (production)
`OTEL_EXPORTER_OTLP_ENDPOINT`	—	OTLP collector endpoint
`OTEL_METRIC_EXPORT_INTERVAL`	`5000` / `60000`	Export interval in ms (5s for console, 60s for otlp)
`OTEL_SERVICE_NAME`	`flatagents`	Service name in metrics