Home

AgentMesh Documentation

AgentMesh is an industrial-grade execution engine for AI agent workflows. Built in C++20 with zero-copy Python interoperability, AgentMesh provides a drop-in replacement for LangGraph runtime execution while delivering up to 1,900x faster node transitions, 95% lower memory footprint, O(1) barrier synchronization, and transaction-safe PostgreSQL crash recovery.

---

At a Glance

import operator
from typing import TypedDict, Annotated, List
from langgraph.graph import StateGraph
import agentmesh_adapter

# 1. Define your standard LangGraph state and workflow
class AgentState(TypedDict):
    messages: Annotated[List[str], operator.add]
    summary: str

builder = StateGraph(AgentState)
builder.add_node("researcher", lambda s: {"messages": ["Found market trend A"]})
builder.add_node("analyst", lambda s: {"messages": ["Synthesized trend A"], "summary": "Bullish"})
builder.add_edge("__start__", "researcher")
builder.add_edge("researcher", "analyst")
builder.add_edge("analyst", "__end__")

# 2. Compile directly onto bare-metal C++20 Engine
app = agentmesh_adapter.compile(builder)

# 3. Execute with standard invoke / ainvoke / stream API
result = app.invoke({"messages": [], "summary": ""})
print(result)

---

Key Performance Guarantees

Metric	LangGraph Native	AgentMesh Engine	Advantage
In-Memory Hop Latency	3.80 ms / hop	0.002 ms / hop	1,900x faster
Diamond Barrier Join	15.60 ms	0.012 ms	1,300x faster
Memory Footprint (100 graphs)	240.0 MB	12.0 MB	95% memory reduction
Crash Recovery Time	Process loss / Unrecoverable	4.8 ms	Zero in-flight state loss
Parallel Fan-Out (N=50)	197.88 ms	187.97 ms	t = 9.78, p < 0.001 (90% Win Rate)
P95 Tail Latency (Cloud Groq)	8,518.31 ms	7,628.35 ms	+889.96 ms (+10.4% lower jitter)

---

Core Capabilities

1. Full LangGraph API Parity: Drop-in support for TypedDict Annotated reducers, Send() dynamic map-reduce, Command(goto=..., update=...), multi-target conditional fan-out, subgraphs, and async streaming.
2. O(1) Atomic Barrier Synchronization: Eliminates polling loops and locks when joining parallel execution branches.
3. Socket-Level Concurrency: Dispatches parallel LLM API calls with zero GIL contention across worker threads.
4. Dynamic Unrolling: Handles cyclic reflection loops and dynamic runtime spawns without pre-compiling execution traces.
5. ACID Crash Recovery: PostgreSQL WAL snapshots restore interrupted workflows in under 5ms with deterministic state resumption.

---

Documentation Index

• Getting Started: Installation | Quickstart | Core Concepts
• How-To Guides: Compile a Graph | Annotated Reducers | Parallel Fanout | Send() Map-Reduce | Command() Routing | Nested Subgraphs | Async & Streaming | Checkpointing & State | Financial Swarm Example
• Architecture: System Overview | C++ Engine Internals | O(1) Scheduler | Dual-Tier Graph | Persistence & WAL | Zero-Copy Pybind Bridge | SOLID Design Principles
• API Reference: agentmesh_adapter | agentmesh_core | LangGraph Parity Matrix
• Benchmarks: Methodology | Results Matrix | Reproducing Results
• Contributing: Development Setup | Adding Tests\n