What tools do I need to build an AI software factory?

Summary

• An AI software factory is a repeatable development system in which agents perform defined work and pass their results through checks and review. People decide what work matters and how much authority the system gets.
• You need six kinds of tools: coding agents, orchestration, isolated execution, verification and review, governance, and observability. A coding agent alone is not a factory.
• On Databricks, Genie Code covers data and AI work in the workspace, and Unity Gateway governs external coding agents such as Claude Code and Codex with model access, rate limits, budgets, and usage tracking.
• Omnigent (Beta) and Databricks Sandbox (Beta) cover multi-agent orchestration and isolated execution. Your existing CI stays in place as the evidence behind each accept or reject decision.
• Start with one narrow task and a human acceptance step, then add layers only when the simpler setup falls short.

What tools do I need to build an AI software factory?

You need six kinds of tools: coding agents that do the work, orchestration that routes tasks, an isolated place to run agent-written code, verification and review gates, governance over models and tools, and observability. A coding agent alone is not a factory. The factory is the process around it: how tasks arrive, how results get checked, and who accepts them.

What is an AI software factory?

This page uses a working definition: a repeatable development system in which agents perform defined work and pass their results through checks and review. People decide what work matters and how much authority the system gets. Keep your existing CI as evidence, and end the automated path at a reviewable proposal until you trust it further.

Which layers does a factory need?

Layer	What it does	Databricks option	Other options
Coding agents	Edit code, run commands, propose changes	Genie Code for data and AI work; external agents such as Claude Code and Codex through Unity Gateway	Any coding agent your team approves
Orchestration	Assigns tasks, combines agents, hands work between them	Omnigent (Beta on Databricks)	LangGraph for custom stateful workflows, or a script for a first version
Isolated execution	Runs agent-written code away from laptops and production	Databricks Sandbox (Beta)	A container or VM sandbox you already operate
Verification and review	Runs tests and checks, then asks a person to accept or reject	Your existing CI; MLflow evaluation for AI features	Your code review process
Governance	Controls which models, tools, and data agents can reach, and how much they spend	Unity Gateway with Unity Catalog permissions	A gateway and policy layer you run yourself
Observability	Shows what each agent did, what it cost, and where it failed	MLflow Tracing and Unity Gateway usage tracking	OpenTelemetry-based tooling

How does Databricks approach each layer?

Coding agents. Genie Code is the AI coding and data assistant for developers in the Databricks workspace. It generates and runs code, builds pipelines and AI/BI dashboards, and debugs errors under your Unity Catalog permissions. Databricks says it is not intended to be a general-purpose programming assistant, so a factory that builds general software also needs an external coding agent. Unity Gateway connects agents such as Claude Code and Codex to governed model services. Developers sign in with Databricks credentials, so no provider API key sits on their machines.
Orchestration and isolation. Omnigent is an open-source meta-harness: a common layer over Claude Code, Codex, Cursor, Pi, and agents you write yourself, with policies and sandboxing. The open-source project is labeled alpha, and the managed version on Databricks is in Beta. Databricks Sandbox (Beta) runs untrusted code in an isolated environment and lists running an agent harness and executing AI-written code as intended uses. Its network egress is currently unrestricted and not configurable, so do not treat it as a network boundary.
Governance and observability. Unity Gateway is a single control plane for model inference, tools, and skills. You grant EXECUTE on a model service to decide who can call it, set request or token rate limits per user, and set budgets that alert or block. Usage lands in the system.ai_gateway.usage system table, and the built-in usage dashboard has a Coding Agents tab with sessions, commits, and lines of code. MCP servers registered in Unity Catalog are governed with the same privileges, and service policies (Beta) add guardrails on requests and responses. For AI features you ship, MLflow Tracing records each step and scorers evaluate the traces.
If the factory ships AI agents. Databricks recommends the Agent Bricks CLI for new agents. It deploys them to Agent Runtime on Databricks Apps, with MLflow tracing configured at deploy time. Agents on Model Serving endpoints are labeled legacy.

FAQs

Is an AI software factory the same as a coding agent?

No. A coding agent performs development tasks with the tools and context it receives. A factory also defines how tasks arrive, how results are checked, and who makes the acceptance decision.

Where does MCP fit?

MCP is an open standard for connecting AI applications to external systems such as data sources and tools. It is how agents reach tools, not an orchestration layer. Unity Gateway governs MCP servers registered as Unity Catalog MCP services.

Do I need every layer on day one?

No. Find the simplest setup that works and add complexity only when it improves outcomes. A script can coordinate a first workflow, and automatic merging is optional.

When do I need Omnigent or LangGraph?

When a script stops being enough. Omnigent fits teams that want to swap or combine coding agents under shared policies. LangGraph fits teams writing their own long-running, stateful agent workflows in code.