Framework Python open-core pour orchestrer des agents IA via un système Kanban event-driven
Project description
AgentFlow
Framework Python open-core pour orchestrer des agents IA spécialisés via un système Kanban event-driven.
Pas de LangChain. Pas de CrewAI. Un Board Kanban comme machine à états, un Event Bus qui découple tout, et des agents qui exécutent — le tout testable sans LLM.
Fonctionnalités
- 5 providers LLM : Claude, OpenAI, Gemini, Mistral, Ollama (local/offline)
- Orchestration Kanban : flux tiré, WIP limits, event bus pub/sub
- Policies pluggables : MaxCost, Timeout, Deadlock, MaxBlocked
- Parallelisme : builders en parallèle via
asyncio.gather - Observabilité : logs structurés JSON, 9 métriques Kanban, cost tracking
- CLI riche : commandes
run,report,list,status— affichage rich ou texte - HTTP entrypoint : API REST FastAPI (optionnel)
- LLM-agnostique : zéro dépendance structurante (pas de SDK LLM)
Installation
# Minimal
pip install agentflow
# Avec affichage CLI enrichi (tableaux colorés)
pip install "agentflow[rich]"
# Avec entrypoint HTTP
pip install "agentflow[http]"
# Tout inclus
pip install "agentflow[all]"
# Depuis les sources
git clone https://github.com/ppaino/agentflow
cd agentflow
pip install -e ".[dev]"
Démarrage rapide
CLI
# Claude (défaut)
export ANTHROPIC_API_KEY=sk-ant-...
agentflow run "Créer une API REST CRUD pour gérer des produits"
# Gemini
export GOOGLE_API_KEY=AIza...
agentflow run "Implémenter un système de cache Redis" --provider gemini --model gemini-1.5-flash
# Mistral
export MISTRAL_API_KEY=...
agentflow run "Refactoriser le module auth" --provider mistral --model mistral-small-latest
# Ollama (local, zéro API key)
# Prérequis : ollama serve && ollama pull llama3
agentflow run "Écrire des tests unitaires pour auth.py" --provider ollama --model llama3
# Dry-run (décomposition uniquement, sans exécution)
agentflow run "Construire un pipeline CI/CD" --dry
# Avec budget et parallélisme
agentflow run "Migrer la base de données" --max-cost 1.50 --parallel
# Lister les sessions passées
agentflow list
# Voir le statut d'une session
agentflow status ./agentflow_sessions/abc123.json
# Rapport complet
agentflow report ./agentflow_sessions/abc123.json
Python API
import asyncio
from agentflow.agents.architect import ArchitectAgent
from agentflow.agents.builder import BuilderAgent
from agentflow.agents.registry import AgentRegistry
from agentflow.agents.reviewer import ReviewerAgent
from agentflow.core.board import Board, BoardConfig, WIPLimits
from agentflow.core.events import EventBus
from agentflow.core.policies import MaxCostPolicy, PolicyEngine, TimeoutPolicy
from agentflow.llm.claude import ClaudeProvider
from agentflow.orchestrator import Orchestrator
async def main():
llm = ClaudeProvider(api_key="sk-ant-...", model="claude-haiku-4-5-20251001")
# 1. Décomposer le brief en cards
architect = ArchitectAgent("architect-01", llm)
cards = await architect.decompose("Créer une API REST pour gérer des utilisateurs")
# 2. Créer le board Kanban
board = Board(
brief="Créer une API REST pour gérer des utilisateurs",
config=BoardConfig(wip_limits=WIPLimits(in_progress=3, review=2)),
)
for card in cards:
board.add_card(card)
# 3. Monter les agents
registry = AgentRegistry()
registry.register(BuilderAgent("builder-01", llm))
registry.register(ReviewerAgent("reviewer-01", llm))
# 4. Policies de sécurité
engine = PolicyEngine()
engine.add_rule(MaxCostPolicy(max_cost_usd=2.0))
engine.add_rule(TimeoutPolicy(max_duration_seconds=300))
# 5. Orchestrer
orch = Orchestrator(board, registry, EventBus(), policies=engine, parallel=True)
final_board = await orch.run()
print(f"Done: {final_board.metrics['cards_done']}/{final_board.metrics['cards_total']}")
asyncio.run(main())
Changer de provider
# OpenAI
from agentflow.llm.openai import OpenAIProvider
llm = OpenAIProvider(api_key="sk-...", model="gpt-4o-mini")
# Gemini
from agentflow.llm.gemini import GeminiProvider
llm = GeminiProvider(api_key="AIza...", model="gemini-1.5-flash")
# Mistral
from agentflow.llm.mistral import MistralProvider
llm = MistralProvider(api_key="...", model="mistral-small-latest")
# Ollama (local)
from agentflow.llm.ollama import OllamaProvider
llm = OllamaProvider(model="llama3")
# Ollama distant
llm = OllamaProvider(model="mistral", host="http://192.168.1.10:11434")
Configuration via variables d'environnement
# Clés API
ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
GOOGLE_API_KEY=AIza...
MISTRAL_API_KEY=...
OLLAMA_HOST=http://localhost:11434 # optionnel
# Session
AGENTFLOW_PROVIDER=gemini # claude | openai | gemini | mistral | ollama
AGENTFLOW_MODEL=gemini-1.5-flash
AGENTFLOW_MAX_COST_USD=5.0
AGENTFLOW_TIMEOUT=600
AGENTFLOW_PARALLEL=true # builders en parallèle
# Kanban
AGENTFLOW_WIP_PROGRESS=3
AGENTFLOW_WIP_REVIEW=2
AGENTFLOW_MAX_RETRIES=3
# Output
AGENTFLOW_OUTPUT_DIR=./sessions
AGENTFLOW_LOG_LEVEL=INFO
AGENTFLOW_LOG_JSON=true
Entrypoint HTTP (FastAPI)
pip install "agentflow[http]"
uvicorn agentflow.http.app:app --reload
Endpoints disponibles :
| Méthode | Endpoint | Description |
|---|---|---|
POST |
/run |
Lance une session (async, retourne session_id) |
GET |
/sessions |
Liste les sessions |
GET |
/sessions/{id} |
Détail d'une session |
GET |
/health |
Healthcheck |
# Lancer une session via HTTP
curl -X POST http://localhost:8000/run \
-H "Content-Type: application/json" \
-d '{"brief": "Créer une API REST", "provider": "ollama", "model": "llama3"}'
Architecture
Entry Points (CLI, Python API, HTTP)
↓
Orchestrateur Kanban (Python pur — ZÉRO appel LLM)
→ Board (machine à états) + Event Bus (pub/sub) + Policy Engine
↓
Agent Layer
→ Architect (décompose) + Builder×N (exécute) + Reviewer (valide)
↓
LLM Abstraction
→ Claude | OpenAI | Gemini | Mistral | Ollama
↓
Telemetry → Structured logs, métriques Kanban, cost tracking
Transitions d'état Kanban
BACKLOG → IN_PROGRESS → REVIEW → DONE
↓ ↓
BLOCKED IN_PROGRESS (retry)
↓
BACKLOG (unblock)
Métriques Kanban
AgentFlow exporte 9 métriques par session :
| Métrique | Description |
|---|---|
throughput |
Cards/heure |
avg_cycle_time |
Temps moyen BACKLOG → DONE (s) |
avg_lead_time |
Temps de vie total d'une card (s) |
wip_peak |
WIP maximum observé |
block_rate |
% cards bloquées |
review_reject_rate |
% cards rejetées en review |
first_pass_yield |
% approuvées au premier passage |
total_cost_usd |
Coût total USD |
token_efficiency |
Tokens output / tokens input |
Tests
# Tests unitaires (zéro appel LLM)
pytest tests/unit/
# Tous les tests
pytest --cov=agentflow --cov-report=term-missing
# Tests d'intégration réels (nécessite clés API)
pytest -m slow
Providers LLM
| Provider | Classe | Modèle défaut | Clé requise |
|---|---|---|---|
| Anthropic Claude | ClaudeProvider |
claude-haiku-4-5-20251001 |
ANTHROPIC_API_KEY |
| OpenAI | OpenAIProvider |
gpt-4o-mini |
OPENAI_API_KEY |
| Google Gemini | GeminiProvider |
gemini-1.5-flash |
GOOGLE_API_KEY |
| Mistral AI | MistralProvider |
mistral-small-latest |
MISTRAL_API_KEY |
| Ollama (local) | OllamaProvider |
llama3 |
aucune |
Licence
MIT
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file agentflow_ppaino-0.1.0.tar.gz.
File metadata
- Download URL: agentflow_ppaino-0.1.0.tar.gz
- Upload date:
- Size: 71.7 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.12.3
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
e778e98ba9cde3eb46d653899fd59472bdd271b08a25f78e6fe6670effa38941
|
|
| MD5 |
6bb1188841307d4643e8c6c4d3fcb57f
|
|
| BLAKE2b-256 |
3184ec271800da5e09f3539fc4471a5160878c2c7df92a455a7140d322060811
|
File details
Details for the file agentflow_ppaino-0.1.0-py3-none-any.whl.
File metadata
- Download URL: agentflow_ppaino-0.1.0-py3-none-any.whl
- Upload date:
- Size: 94.8 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.12.3
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
8e77351b20253c044345db2c80dd7267a96742ec931217d6f3317d3bd7fe29d8
|
|
| MD5 |
dd890bd2bfbd38839448a9dcfed80b19
|
|
| BLAKE2b-256 |
fbd7ab2298ba262ead834007db4b3404648dd8d6f6247125bb43b8bc51fc9313
|