Skip to content

Executive summary

Target

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

The problem

Organizations are adding AI agents to their work faster than they can govern them. Each tool brings its own identity, its own memory, its own idea of what an approval is; what the agents learn leaves with the vendor; and nobody can say afterwards what ran, with which data, under whose authority, at what cost. Software delivery shows it first: generated code without provenance, deployments without gates, products that nobody operates.

The answer

The Factory is the organization's AI operating and control platform. It gives every piece of agent work the same four things: a scope (who, which project, which environment, which data), gates (every side effect has a class, and the gated classes need a human decision bound to the exact version), evidence (receipts, timelines, audit) and ownership (published assets, knowledge, memory and execution history stay with the organization while models are swapped). The first thing built on it is the Software Factory: products generated from ConnectSoft's own templates, delivered through governed pipelines and operated by ConnectSoft as a managed service.

flowchart LR
    P[People · seven roles] -->|submit · review · operate| C[Control plane · runs · approvals · assets]
    C --> A[Agents · skills · workflows]
    A -->|classed tools under policy| T[Tools · MCP · sandboxes]
    A --> M[Model gateway · profiles · receipts]
    A --> K[Knowledge · memory · artifacts]
    C --> E[Evidence · audit · evaluation · usage]
    S[Software Factory pack] --> G[Generated products · delivered · operated]
    C --- S
Hold "Alt" / "Option" to enable pan & zoom

What the design commits to

  • Nine planes, sixteen platforms, forty-two bounded contexts, about 180 logical services, every one specified to the table and the endpoint, generated from one architecture model (platform map).
  • Modern agent protocols: MCP for tools, A2A for agents, OpenTelemetry GenAI conventions for observability, workload identity for agents, durable execution (technology radar).
  • Quality gates: no agent, skill, workflow, tool, model profile or template is published without evaluation evidence and no regression (quality gates).
  • Data governance: tenant-selected region, classification, retention, deletion within a window, export in open formats (data governance).
  • A pilot-class service level stated as numbers, not adjectives (NFR baseline).

What exists today

ConnectSoft's estate is real and in production use: the template library (43 templates, about 125 libraries), the identity and authorization services, the SaaS foundations, the messaging and persistence libraries, the AI extensions on the Microsoft Agent Framework, Pulumi programs and pipelines — proven by a production product (DerechHaLev) built from them. The Factory as a platform does not exist yet: no run control, no agent runtime, no gateway, no Studio. The ecosystem composition says component by component what exists, what is extended and what is developed.

How it is operated and sold

ConnectSoft operates generated products as a managed service with per-product objectives, rollouts and incident handling (decision D33). The Factory itself is sold under editions that cover internal budget control, usage-based billing and seat plans, on the same SaaS foundations ConnectSoft ships to its customers (commercial model).

Risks and controls

Risk Control in the design
Agents act beyond authority side-effect classes, policy that only tightens, approvals bound to exact versions with separation of duties
Learning leaks across tenants or into vendors tenant-scoped memory, promotion only with review and consent, model profiles instead of vendor lock-in
Quality erodes as agents change evaluation gates with thresholds and no-regression rules before publication; drift detection in production
The platform is too large to build dependency order without dates, modules of hosts rather than 180 deployables, generated catalogs that keep design and model in step
Nobody can prove what happened receipts, timelines, hash-chained audit with 7-year retention, usage by status