Agent as a Backend
Your agent doesn't need
a backend. It is the backend.
Define capabilities, not endpoints. Agentend handles intent routing, memory hydration, model orchestration, and event streaming — so you ship agent-powered products, not infrastructure.
The Paradigm Shift
What if you didn't need any of this?
Traditional Backend
Agent as a Backend
All that complexity collapses into intent-driven capabilities.
This is all you write
No routes. No controllers. Just capabilities.
from agentend import Agentend, Capability, tool
app = Agentend()
@app.capability("invoice_processing")
class InvoiceProcessor(Capability):
"""Extract, verify, and summarize invoices."""
workers = ["extract", "verify", "summarize"]
def get_domain_context(self, ctx):
return "You process business invoices. Be precise with amounts."
@tool
def save_invoice(self, data: dict) -> str:
db.invoices.insert(data)
return f"Invoice {data['id']} saved"
if __name__ == "__main__":
app.serve()How Agent as a Backend Works
01
User sends an intent
Natural language, not a structured API call. The agent understands what you need.
02
Kernel classifies & routes
Small model, <10ms. Picks the right Capability from the registry.
03
Context Bus hydrates memory
5 tiers, progressive loading. Only pulls what the task actually needs.
04
Workers execute, events stream
Typed workers call models. AG-UI protocol streams results in real-time.
6
Worker Slots
5
Memory Tiers
13
Event Types
18
Modules
Everything you need. Nothing you don't.
18 modules, zero boilerplate.
Fleet
6 typed worker slots with benchmark-backed model defaults
Memory
5-tier context bus with progressive hydration
Events
AG-UI protocol — 13 event types streamed via SSE
Orchestrator
DAG workflows with parallel steps and HITL interrupts
Protocols
AG-UI + MCP + A2A — the protocol triangle
Security
PALADIN 5-layer injection defense, built in
Cache
Dual-layer semantic cache — hash + vector similarity
Budgets
Per-tenant token budget enforcement with hard limits
Plugins
8 hook points, YAML manifests, pip-installable
Observability
OpenTelemetry traces + metrics on all critical paths
Persistence
SQLAlchemy 2.0 async with multi-tenant RLS
Builder
Chat-to-build: create capabilities without code
Architecture
Start building in minutes
Production-ready agent backends. Not another wrapper.