The Lowpass Dispatch

OpenAI is expanding its Codex coding agent from a developer tool into a full-fledged enterprise platform. A new partnership with Dell announced in May 2026 will bring Codex to hybrid and on-premise environments, allowing companies to run coding agents securely on their own infrastructure.

The platform strategy is supported by major enterprise adoptions. Databricks is using GPT-5.5, the model behind the latest Codex, for its enterprise agent workflows. Similarly, Asian tech giant Sea Limited is deploying Codex across its engineering teams to accelerate development. To ensure safe execution, OpenAI has also detailed its work on a secure sandbox for Codex on Windows.

The push for ubiquity extends to individual developers, with a new feature in the ChatGPT mobile app allowing users to monitor and steer Codex tasks remotely. Together, these moves show a clear strategy: make Codex a secure, accessible, and indispensable part of the entire software development lifecycle, from corporate data centers to a developer's phone.

A technology company recently rewrote its separate legacy iPhone and Android applications into a single React Native codebase, a task accomplished primarily with coding agents. The move highlights a new dynamic in software engineering, where the cost and risk of major architectural changes are falling dramatically.

According to a May 14 report, the team's rationale was that if the migration proved to be a mistake, they could "just port back to native" using the same agent-driven process. This sentiment echoes a quote from programmer Mitchell Hashimoto: "Programming languages used to be LOCK IN, and they're increasingly not so."

This shift changes the calculus for technical debt and platform lock-in. Projects previously deemed too expensive or risky, like modernizing legacy systems or switching frameworks, are becoming feasible. For engineering leaders, it means strategic agility can be valued more highly over picking the "perfect" long-term stack from the outset.

Researchers have introduced LivePI, a new benchmark for testing AI agents against indirect prompt injection (IPI). Unlike previous simulated tests, LivePI runs agents in a production-like virtual machine with live, test-controlled access to email, chat, web browsers, and local files, exposing them to attacks embedded in untrusted inputs.

The benchmark covers seven input surfaces and five malicious goals, including data exfiltration and unauthorized code execution. Across five leading models, including GPT-5.3-Codex and Claude Opus 4.6, total attack success rates ranged from 10.7% to 29.6%. The researchers noted that attacks delivered via group-chat messages were "uniformly successful" across all tested models.

The findings underscore the significant security risks in deploying agents that interact with external data sources. LivePI provides a more realistic methodology for evaluating these risks and testing defenses, such as prompt filtering and pre-execution authorization, before deploying agents into production environments.

A survey paper from May 2026 titled "Code as Agent Harness" articulates a fundamental shift in how developers are building AI agents. The paper argues that code is no longer just a target for generation, but is increasingly used as the operational infrastructure for the agent itself.

This framework views code as the medium for agent planning, memory, and tool use. Instead of complex prompt chains, agents can generate and execute code to perform actions, verify results, and interact with their environment. This approach provides a more structured and verifiable way to build complex, long-horizon agentic systems.

The "harness" concept provides a unified lens for understanding disparate trends, from repository-level software engineering agents to GUI automation. It suggests that the future of agent development will rely heavily on robust software engineering principles applied not to the final product, but to the agent's own internal processes.