MeshFlow
  • Cloud PlatformDeployIntegrationsInit CLIWhy MeshFlow
  • DocsLearn
  • CustomersCertificationBlogChangelogCommunity
  • Pricing
Sign inGet started
Pricing
Sign inGet started
MeshFlow

The production-safe standard for agentic AI.
Apache 2.0 — free forever.

Product
  • Pricing
  • Cloud
  • Compare
  • Init CLI
  • Changelog
Docs
  • Quickstart
  • Core concepts
  • Compliance guide
  • Token optimization
  • API reference
Resources
  • Blog
  • Customers
  • Community
  • PyPI
  • GitHub
Company
  • Discord
  • Twitter
  • Security
  • License
MeshFlow © 2026 · Yaya Systems · Apache 2.0
All systems operational
MeshFlowdocs
Search docs...⌘K
v1.13.0 GitHub Discord
Getting startedQuickstartInstallationCore conceptsHarness architectureSkills vs ToolsSandbox mode
GuidesBuilding agentsMulti-agent workflowsModel RoutersCost AnalyticsCompliance guideToken optimization
Claude & AIAdvisor patternThinking budgetsDynamic workflowsContext compactionTool streaming
Governancemeshflow-forensicZero-Trust PoliciesSOC 2 assertionEU AI ActProduction Readiness
IntegrationsOpenAI Agents SDKLangGraphCrewAIMCP ToolsTypeScript SDKOTel / OTLPAutoGen 0.4+
API ReferenceWorkflowAgentCostCapAdvisorAgentCostRegressionGate
Docs/Getting started/Harness architecture

Harness architecture

The AI Agent Harness is the foundational infrastructure that wraps around an LLM to transform it from a stateless reasoning engine into a reliable, autonomous, and production-ready agent. MeshFlow acts as your pre-built Agent Harness with 12 core pillars.

1. Core Execution

The Engine (Execution Loops): Handled by MeshFlow's Team and StateGraph run loops.

Delegation (Topologies): Handled by hierarchical teams and Team(pattern="supervised") where a planner agent spawns worker agents.

Context Management: Handled by Context Compaction strategies.

2. The Capability Layer

Tools (External Actions): The meshflow.tools registry exposes strictly-typed external functions.

Skills (Internal Cognitive Abilities): Semantic functions that teach the agent how to think, via the skills=[] parameter.

MCP (Model Context Protocol): Standardized injection of context and APIs.

3. State & Resilience

Session Persistence: Accomplished via ReplayLedger.

Memory Architecture: Agent memory components, RAG modules, and Cross-Run Learning (CORAL).

Error Recovery: Automatic reflection and circuit breakers on execution timeouts.

4. Governance & Operations

Lifecycle Hooks: Human-in-the-loop (HITL) checkpoints.

Guardrails & Permissions: PII scrubbers and strict CostCaps.

Observability: OpenTelemetry integration and SOC2/EU AI Act audit reporting.

Previous
Core concepts
Next
Skills vs Tools
On this page1. Core Execution2. The Capability Layer3. State & Resilience4. Governance & Operations