Skip to content

Agent Mesh — Deployment

Target

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

Runtime model

Concern Choice
Deployment unit AgentRuntimeWorker, one image from the AI agent host template with the worker hosting shape
Hosting AKS, 2 to 20 replicas, scaled on free attempt slots and inbox lag
Data runtime-sql (attempts, checkpoints, calls, decisions, delegations, loaded releases, workflow runs, timers); artifact-blob for adapter checkpoints and produced artifacts; cache for release manifests
Actors Orleans silo in-process; AttemptGrain, WorkflowRunGrain; reminders for timers and heartbeats
Messaging MessagingModel inbox and outbox; Orleans profile by default
Identity workload identity from the Authorization Server; per-attempt tokens derived from leases
Ingress none from the public internet; GW-03 for Studio reads; gRPC from the Control plane
Egress model gateway, tool executor, Knowledge, Control plane, artifact store; everything else denied

Topology by profile

flowchart TB
    subgraph CS["ConnectSoft cloud, tenant region"]
        CP["ControlPlaneApi"]
        MG["ModelGatewayHost"]
        TE["ToolExecutorHost"]
        KH["KnowledgeHost"]
        W1["AgentRuntimeWorker (saas-default)"]
        SQL1["('runtime-sql')"]
    end
    subgraph Cust["Customer network (hybrid)"]
        W2["AgentRuntimeWorker (hybrid)"]
        SQL2["('runtime-sql')"]
        TE2["ToolExecutorHost (sandboxes near the code)"]
    end
    W1 --> SQL1
    W1 --> CP & MG & TE & KH
    W2 --> SQL2
    W2 -- "private link" --> CP & MG & KH
    W2 --> TE2
Hold "Alt" / "Option" to enable pan & zoom
Profile Worker placement
saas-default in ConnectSoft's AKS in the tenant's region
hybrid inside the customer's network next to repositories and tools; outbound-only connectivity to the Control plane over a private link
on-prem with the rest of the Factory in the customer's cluster; no broker required
hybrid-compute in ConnectSoft's cloud; model and compute pools are the customer's

Scaling

Signal Action
free attempt slots below threshold add replica (capacity is advertised to BC-26 per replica)
inbox lag above threshold add replica; readiness fails on the lagging replica
attempts waiting for decisions no scaling; waiting attempts hold no slot beyond their grain
runtime-sql write latency partition checkpoints by tenant and month; compaction window

Tenant fairness: Execution Capacity places attempts per tenant quota; a replica never takes more than MaxConcurrentAttempts.

Configuration and secrets

Options sections Runtime:Host, Runtime:Adapter, Runtime:Attempts, Runtime:Context, Runtime:Decisions, Workflows:Engine, OrleansMessaging, ConnectSoftIdentityPlatform; reloadable ones from the Config Platform. Secrets in Key Vault through workload identity; sandbox credentials are never configured on the worker, they are brokered per call.

Health, readiness, rollout

/runtime/health liveness and readiness (store, silo, Control plane, gateway, executor, inbox lag). Rolling updates drain a replica: no new dispatches are claimed, running attempts checkpoint and migrate their grains, then the pod stops. An adapter version change is a rollout that marks older checkpoints incompatible; Run Control starts new attempts for them.

NFR class and recovery

pilot-interactive (D30). runtime-sql is restored after the run journal and before the outbox and artifacts; after a restore the engine re-projects from Run Control's journal and attempts whose checkpoints are missing end Uncertain.