US immigration agency wants AI agents to write and test its software
ICE has asked industry how to build an "agentic software factory", with a rule that no AI component may approve its own release.
US Immigration and Customs Enforcement (ICE) is seeking help to build an "agentic software factory", in which AI agents handle much of the planning, coding, testing and review of software, according to a request for information reported by FedScoop.
The agency describes three layers it wants, which it says are "desired future-state capabilities, not existing ICE implementations": an enterprise control plane for governance, infrastructure and cost reporting; an orchestration layer that coordinates agents and sets rules for tool access, escalation to humans and stop conditions; and an independent assurance boundary that tests for cybersecurity, accessibility and mission acceptance before anything reaches production.
Separation of duties for machines
ICE says the factory would expand its existing STELLA platform, short for Secure Technology Environment for Large Language Agents, which went live this summer and already hosts at least a dozen production AI applications. Its stated design rule is that "generation, assurance, and authorization must remain distinguishable and no autonomous component may silently expand its own permissions or approve its own production release".
That principle is borrowed from long-standing financial and security controls: the person who writes code should not be the one who approves it. Applying it to AI agents is a sensible starting point for any agency experimenting with automated development.
Questions the document leaves open
Software factories built by agents promise speed, but they also move risk. Code written faster still needs people who understand it when it fails, and accessibility and security testing must be real gates rather than automated rubber stamps. The request for information is an early market-sounding step, not a contract. How ICE answers those questions in the eventual solicitation will matter more than the architecture diagram, particularly for systems that support enforcement decisions about people.
Newsroom, Online
Stories by the IQGov newsroom are researched and written by AI agents to our editorial rulebook, checked against their sources by a separate fact-checking agent, and approved by the publisher before publication. Every fact links to its source.
