How to build an enterprise-grade AI software factory

Summary

Matar Peles, Port’s Field CTO, explains how to build an enterprise-scale AI software factory: a software delivery process in which AI agents lead work through stages from an idea to a pull request, with human review and approval where needed.

The central argument is that AI has made code generation faster, shifting the bottleneck to the rest of the software development life cycle—such as planning, reviews, approvals, and coordination. In an AI-native SDLC, agents draft and advance work at each stage; people set direction, review outputs, and intervene according to risk.

Main Concepts and Architecture

• Factory lines: Organizations can create workflows for feature delivery, infrastructure provisioning, incident remediation, CI/CD fixes, or security issues such as CVEs. A feature-delivery example runs through intake, planning, design, build, test, deployment, and maintenance.
• Deterministic guardrails: The challenge is not simply connecting agents. Enterprise workflows need predictable steps around agents’ less predictable work. Workflows can include conditions, approvals, and standard automation; not every stage needs an AI agent.
• Unified context layer: Agents need a shared, current view of organizational information, such as services, teams, ownership, repositories, cloud resources, incidents, and issues. In the demo, Port represents this information as a connected data model or graph, accessible to agents through MCP and populated by integrations with existing tools. Peles says a shared context layer can reduce agent costs; he cites Port research suggesting reductions of up to 80% compared with repeatedly gathering context through multiple MCP calls.
• Agent, skill, and MCP registries: A central registry provides visibility into which agents and tools are in use, who owns them, what they can access, where they run, and what they cost. The presenter describes certification and readiness scoring for agents and skills, along with self-service provisioning through approved “golden paths.”
• Governance and oversight: Enterprises need access controls for people and agents, clear workflow ownership, audit trails, budget limits, and approval rules. Human involvement can be determined by risk, service tier, or the potential impact of a change.
• Measurement and improvement: Dashboards can track agent costs, time, budgets, rework, workflow status, and the quality of outputs such as specifications. The presenter advocates measuring from the start and using results to improve workflows over time.

Demo and Practical Guidance

The Port demo walks through an example factory based on Anthropic’s AI-native SDLC approach. It shows a context graph, agents querying that context, an agent registry, skill certification workflows, and a feature-delivery workflow. A control-room dashboard tracks spending and workflow performance. The presenter emphasizes that this is an example, not a universal process: teams should adapt workflows to their products and requirements.

Peles argues that one factory line is unlikely to suit an entire company. Platform teams should provide shared infrastructure, governance, and standards while enabling individual teams to build or adapt their own workflows.

In a Q&A, he distinguishes Port as an organizational and end-to-end workflow platform from LangChain or LangGraph, which he characterizes as tools more focused on the agent harness and agent-level architecture. He suggests they can be used together.

Guides, Tutorials, and Reviews

• Technical walkthrough and live product demo: The session is primarily an architecture explanation followed by a Port demonstration; it is not an independent product review.
• Referenced material: Peles points to Anthropic’s AI-native SDLC playbook and mentions a previous webinar covering Anthropic’s implementation in more depth. He also notes that OpenAI and other companies have published descriptions of their software-factory approaches.

Main Speakers and Sources

• Matar Peles, Field CTO at Port — main presenter and source of the architecture guidance and product demonstration.
• Platform Engineering webinar host — introduces the session, moderates audience questions, and closes the webinar.
• Referenced sources: Anthropic’s AI-native SDLC playbook; examples of software-factory architectures from Anthropic and OpenAI, with Uber also mentioned.