Skip to content

Agent Mesh — Overview

Target

This page describes intended behaviour or structure. No implementation is claimed unless a section is explicitly marked Current implementation or Evidence.

Rewritten 2026-10-09 (design plan P2)

This page replaces the June-2026 overview that described eleven ConnectSoft.Factory.AgentMesh.* services including registries, a model router and a tool adapter layer. Those responsibilities now live with their owners: publication in the Control plane (BC-04), model routing in Model & Compute (BC-07), tool execution in Tools & MCP (BC-08), capacity in Runtime Cloud (BC-26). The Agent Mesh keeps what is its own: executing attempts and driving workflows. The old page is archived under catalogs 2026-06.

The Agent Mesh is where work is done and nowhere else: one attempt of one step, under one lease, against one immutable snapshot, with every external effect routed through a governed executor and every resumable point checkpointed. Dynamic agent reasoning (the plan-call-observe-verify loop) and deterministic orchestration (DAG workflows) coexist in the same worker, with Run Control as the single authority over both.
Platform PF-02 Agent Mesh (Agent & Orchestration runtime)
Bounded contexts BC-06, BC-38
Services 6 (1 hosts, 5 modules, 0 external, 0 later) — 3 modelled
Gateways —
Status target

Purpose

  • Execute, never decide. The runtime claims leases, runs attempts and reports outcomes; the Control plane accepts, dispatches, approves and records.
  • Survive anything. Process restarts, node loss, provider outages, tool failures and approval delays never lose a run: checkpoints before every gated call, durable timers, lease reconciliation and Uncertain outcomes instead of guesses (ADR-0024, Proposed).
  • Coexist. Agents reason dynamically inside a step; workflows orchestrate deterministically across steps with conditions, retries, compensation, timers and human gates (WorkflowEngineService).
  • Stay replaceable. The Microsoft Agent Framework is the execution adapter behind a contract with versioned checkpoints; a second adapter passes the same suite.

Bounded contexts and services

Context Owns Services (granularity)
BC-06 Execution Attempt execution, checkpoints, call records, decision requests, sub-agent delegations, loaded releases AgentRuntimeService (module); AgentExecutionService, AgentValidationService, AgentCorrectionService merged
BC-38 Workflow Definitions & Engine Workflow definition versions, workflow run projections, step runs, timers, compensation WorkflowEngineService (module)
Host Composition, execution adapter, workload identity, capacity advertisement AgentRuntimeWorker (standalone)
Service Context Granularity Status
AgentCorrectionService BC-06 module of host unmodelled
AgentExecutionService BC-06 module of host unmodelled
AgentRuntimeService BC-06 module of host target
AgentRuntimeWorker BC-06 standalone host target
AgentValidationService BC-06 module of host unmodelled
WorkflowEngineService BC-38 module of host target

Role in the Factory

flowchart LR
    CP["Control Plane<br/>Run Control, Asset Publication"] -- "AttemptDispatched, ApprovalDecided, releases" --> Mesh["Agent Mesh<br/>AgentRuntimeWorker"]
    Mesh -- "ReportAttemptOutcome, RequestApproval, StepReady" --> CP
    Mesh -- "model calls with receipts" --> MC["Model & Compute"]
    Mesh -- "tool and MCP calls with policy decision ids" --> Tools["Tools & MCP"]
    Mesh -- "context snapshots, artifacts" --> KM["Knowledge & Memory"]
    Mesh -- "capacity, sandboxes" --> RC["Runtime Cloud<br/>Execution Capacity"]
    Mesh -- "spans, timeline events" --> Obs["Observability"]
Hold "Alt" / "Option" to enable pan & zoom

The execution loop

flowchart TB
    D["AttemptDispatched<br/>lease + fencing token"] --> R["Resolve and load release<br/>verify hashes, adapter compatibility"]
    R --> C["Build context snapshot<br/>Knowledge, token budget"]
    C --> P["Plan"]
    P --> M["Model call<br/>profile, receipt"]
    M --> T{"Tool call?"}
    T -- "read / sandbox" --> X["Execute via Tools & MCP"]
    T -- "gated class" --> A["Policy decision, approval bound to proposal hash, checkpoint"]
    A --> X
    X --> O["Observe, update state"]
    O --> V{"Verified?"}
    V -- no, corrections left --> P
    V -- yes --> Rep["Report one outcome<br/>hash + fencing token"]
    V -- corrections exhausted --> Rep
Hold "Alt" / "Option" to enable pan & zoom

Component view

flowchart LR
    subgraph P["Agent Mesh"]
        AgentRuntimeWorker["AgentRuntimeWorker"]
    end
    P --> RunControlService["RunControlService"]
    P --> AgentRegistryService["AgentRegistryService"]
    P --> ModelRouterService["ModelRouterService"]
    P --> ToolAdapterService["ToolAdapterService"]
    P --> ContextBuilderService["ContextBuilderService"]
    P --> ArtifactService["ArtifactService"]
    P --> PolicyEngineService["PolicyEngineService"]
    P --> AgentPoolManager["AgentPoolManager"]
    P --> ControlPlaneApi["ControlPlaneApi"]
    P --> ModelGatewayHost["ModelGatewayHost"]
    P --> ToolExecutorHost["ToolExecutorHost"]
    P --> KnowledgeHost["KnowledgeHost"]
    P --> RuntimeCloudHost["RuntimeCloudHost"]
    P --> ext_identity_platform["ext-identity-platform"]
Hold "Alt" / "Option" to enable pan & zoom

Built from the ConnectSoft estate

Component Use here Status
AI agent host template (ConnectSoft.AI.SoftwareFactory.AgentTemplate) AgentRuntimeWorker host extend
Worker template hosting shape, health, scheduling exists
ConnectSoft.Extensions.AI and .AgentsFramework Microsoft Agent Framework adapter, keyed model clients extend
ConnectSoft.Extensions.AI.Factory lifecycle skeleton (accept, prompt, run, validate, retry) to become the attempt loop extend
ConnectSoft.Extensions.ModelContextProtocol MCP client side of tool calls extend
MessagingModel Orleans adapter inbox, outbox, grains, reminders exists
Observability libraries OpenTelemetry GenAI spans exists

Modernization (2 to 3 year horizon)

  • OpenTelemetry GenAI semantic conventions on every model and tool span, joined to the run trace.
  • Structured outputs and tool-call reliability as evaluation gates (D32) rather than runtime heuristics.
  • Agent memory and governed learning: the runtime proposes, Knowledge promotes after review (D32, D33).
  • Computer-use and browser agents as sandboxed tools of the Tools & MCP platform, never as runtime privileges.
  • A2A: supervisor attempts may delegate to partner agents as tools of class external (ADR-0023).

Current implementation

ConnectSoft.Extensions.AI.Factory (FactoryAgentBase lifecycle without persistence, lease or cancellation), ConnectSoft.Extensions.AI.AgentsFramework (keyed agents, explicit tools, RAG, file skills, DevUI) and the agent host template (direct RunAsync entry paths, credit-counter grain, logging-only jobs, saga scaffold) exist; no attempt, checkpoint, lease, workflow engine or release loader exists. The Platform repository holds six placeholder workflow definitions that nothing executes.

Final-state summary

One worker host, two modules, one execution adapter behind a contract, Orleans grains for durability, the Control plane for truth. Every attempt is resumable or honestly Uncertain; every workflow is data validated at publication; every effect is a governed call with a receipt.