Skip to content

Agent Mesh Platform — Overview (June 2026)

Archived 2026-10-09

This June-2026 narrative page was superseded by the rewritten page at planes/agent-runtime/overview.md (design plan P2). Registries moved to the Control plane (BC-04), model routing to Model & Compute (BC-07), tool adapters to Tools & MCP (BC-08) and pooling to Runtime Cloud (BC-26); validation and correction became phases of the attempt loop. It is kept for provenance.

Historical — archived

This page is kept for provenance. Its dates, targets and technology statements are not current direction. See the linked replacement page.

The Agent Mesh was described as the platform that hosts and operates ConnectSoft's autonomous agents: it claims agent tasks, loads governed context packages from the Knowledge Platform, executes skills on the Microsoft Agent Framework runtime, routes model calls, invokes tools, validates outputs, self-corrects, and emits every event in the canonical envelope.

Purpose (June 2026)

Make autonomous software production safe, reusable, observable and traceable at multi-tenant scale: a single governed runtime in which any agent role is interchangeable, every execution is grounded in retrieved knowledge, and every artifact traces back to the task, agent, skill, model and context that produced it.

Role in the Factory (June 2026)

flowchart LR
    Intent["Business Intent"] --> CP["Control Plane"]
    CP -->|"assigns AgentTask"| Mesh["Agent Mesh"]
    Mesh -->|"requests ContextPackage"| Knowledge["Knowledge Platform"]
    Knowledge -->|"grounded context"| Mesh
    Mesh -->|"model calls"| Integration["Integration Platform"]
    Mesh -->|"produces artifacts"| Templates["Template Library Platform"]
    Mesh -->|"emits events"| Observability["Observability & Feedback"]
    Mesh -->|"enforced by"| Governance["Governance, Security & Compliance"]
Hold "Alt" / "Option" to enable pan & zoom

Core Responsibilities (June 2026)

Responsibility Description
Agent hosting Run agents on the Microsoft Agent Framework runtime via ConnectSoft.Extensions.AI.AgentsFramework, with managed lifecycle, pooling, and health.
Task processing Claim, execute, validate, correct, and complete agent tasks against the task contract.
Context grounding Request and consume context packages from the Knowledge Context Builder.
Skill execution Resolve and run versioned skills with explicit input/output contracts.
Model routing Select and invoke the right model provider through the ModelRouterService.
Tool adaptation Expose governed tools to agents over the Model Context Protocol (MCP).
Validation & correction Validate every output against schema, naming, dependency, and policy rules; self-correct within bounded attempts.
Event emission Emit lifecycle, domain, and integration events in the canonical envelope.
Observability Capture telemetry, traces, and health for every model and tool invocation.

Platform Building Blocks (June 2026)

Service Responsibility 2026-10 owner
AgentRuntimeService Hosts agents; orchestrates a single execution BC-06 (kept)
AgentRegistryService AgentDefinition / AgentVersion BC-04 Control plane
SkillRegistryService SkillDefinition / SkillVersion BC-04 Control plane
ModelRouterService Policy-driven model provider selection BC-07 Model & Compute
ToolAdapterService MCP tool adapters BC-08 Tools & MCP
AgentTaskService AgentTask intake and lifecycle BC-03 Run Control
AgentExecutionService AgentExecution / SkillExecution merged into AgentRuntimeService
AgentValidationService ValidationResult merged into AgentRuntimeService
AgentCorrectionService CorrectionAttempt merged into AgentRuntimeService
AgentTelemetryService ModelInvocation / ToolInvocation telemetry BC-21 Observability
AgentPoolManager AgentHealthStatus, pooling, warm capacity BC-26 Runtime Cloud

Final-State Summary (June 2026)

A multi-tenant, policy-governed, fully observable runtime built on the Microsoft Agent Framework: every agent registered and versioned, every skill reusable, every execution grounded in retrieved knowledge, every output validated and self-corrected, every action emitted as a traceable event. The 2026-10-09 design keeps that intent and narrows the platform to attempt execution and the workflow engine under Run Control's authority.