AgentFlow

AI-Native Communication Intelligence for Multi-Agent Systems

AgentFlow is a multi-agent orchestration platform with an intelligent communication layer that reduces LLM API costs by 40-70% through intent classification, predictive caching, delta encoding, and state deduplication.

Quick Start

pip install agentflow

Define and run a pipeline

from agentflow import Agent, Pipeline, Orchestrator

agent_a = Agent("researcher", tier=1)
agent_b = Agent("analyst", tier=1, depends_on=["researcher"])
agent_c = Agent("writer", tier=2, depends_on=["analyst"])

pipeline = Pipeline()
pipeline.add_agent(agent_a)
pipeline.add_agent(agent_b)
pipeline.add_agent(agent_c)

orchestrator = Orchestrator(pipeline)
results = orchestrator.run("Analyse steel supply chain risks")

Classify agent intents

from agentflow.comms.intent_classifier import IntentClassifier

classifier = IntentClassifier()
classifier.load("data/intent_model.pkl")

result = classifier.classify("what is the current throughput")
print(result)  # {"intent": "query", "confidence": {"query": 0.77, ...}}

Start the API server

# In-memory (development)
AGENTFLOW_ADMIN_KEY=your-key uvicorn agentflow.api.server:create_app --factory --host 0.0.0.0 --port 8100

# With persistence (production)
AGENTFLOW_ADMIN_KEY=your-key AGENTFLOW_DATABASE_URL=sqlite+aiosqlite:///agentflow.db uvicorn agentflow.api.db_app:create_persistent_app --factory --host 0.0.0.0 --port 8100

API authentication

All endpoints (except /health) require an API key header:

curl -H "X-API-Key: your-key" http://localhost:8100/events/summary

Architecture

AgentFlow has four layers:

Layer 1 — Orchestration: Wave-based parallel execution, dependency resolution, fault tolerance, multi-provider LLM routing with circuit breaker and key rotation.

Layer 2 — Communication Intelligence: Intent classification (TF-IDF + logistic regression, 99.96% accuracy), delta encoding (RFC 6902 JSON Patch + msgpack), state deduplication (content-addressed SHA-256), predictive caching (structural template learning).

Layer 3 — Observability: AgentPulse per-agent cost attribution, compression metrics, intent distribution analytics, Streamlit dashboard.

Layer 4 — Enterprise: RBAC (admin/operator/viewer), rate limiting, output filtering, CompressProxy for legacy integration, Docker + Helm deployment.

Database

AgentFlow uses SQLAlchemy 2.0 async with Alembic migrations. SQLite for development, PostgreSQL for production.

# Run migrations
alembic upgrade head

# Switch to PostgreSQL
export AGENTFLOW_DATABASE_URL=postgresql+asyncpg://user:pass@host/agentflow
alembic upgrade head

API Endpoints

Security

• API key authentication (SHA-256 hashed, never stored in plaintext)
• Role-based access control (admin, operator, viewer)
• Per-endpoint permission mapping
• Rate limiting with burst detection
• Output filtering (prompt leak and PII detection)
• Input sanitisation
• Canary token detection
• Request signing
• AES-GCM encryption at rest
• Row-level tenant isolation
• Audit trail on all operations
• Soft deletes (GDPR-friendly)

Configuration

Environment variables:

Installation with extras

pip install agentflow[postgres]    # PostgreSQL support
pip install agentflow[dashboard]   # Streamlit dashboard
pip install agentflow[dev]         # Development tools

Pricing

AgentFlow is a paid product with three tiers. There is no free tier — the orchestrator is the core USP.

Starter (69/month): 100K messages, 20 agents, sequential execution + intent classification. Pro (199/month): 1M messages, 50 agents, parallel wave execution + full compression. Enterprise (599/month): Unlimited, self-hosted option, SSO/RBAC, SLA.

License

Proprietary — Sedge5 Technology Solutions