How AI Agents Are Redefining Software Development: The New Agent‑Native Paradigm

Three independent teams—OpenAI, StrongDM, and the author’s own firm—discovered the same workflow for building software with AI agents, despite no prior coordination. Their experiments show a fundamental shift: the core engineering challenge moves from writing complex logic to constructing environments where agents can reliably generate and validate code.

OpenAI’s Practice

Starting from an empty Git repository in August 2025, OpenAI’s team produced roughly one million lines of code, merged about 1,500 pull requests, and grew from three to seven engineers over five months, achieving an estimated ten‑fold speedup over manual coding. Every line—application logic, tests, CI configuration, documentation, internal tools—was written by Codex, with humans never directly contributing code.

“Our primary work became empowering the agent to do effective work.”

When tasks fail, the focus shifts from improving prompts to identifying missing capabilities or context for the agent, building feedback loops, structured documentation, and runtime environments. A single Codex run can work continuously for six hours, often while the team sleeps.

StrongDM’s Practice

Since July 2025, a three‑person team at StrongDM built a “Software Factory” where agents write and review all code without human intervention. They treat generated code like model weights—opaque and verified only through external behavior, using a “Digital Twin Universe” that clones APIs from Okta, Jira, Slack, and Google Docs to allow thousands of hourly integration scenarios without rate limits or costs.

Testing scenarios are stored outside the codebase as a “Holdout Set” to prevent agents from gaming the tests. Their coding agent, Attractor, consists of three Markdown specification files that any modern coding agent can use to self‑assemble.

Common Patterns Identified

“Boring” tech wins : Both teams chose composable, stable, well‑documented tools (Go, Rust, TypeScript) because agents perform better with widely represented, stable APIs.

Repository is the only truth : Knowledge must reside in version‑controlled files; agents cannot see Slack, brains, or Google Docs. StrongDM uses a tiny AGENTS.md (≈100 lines) as an index to other structured docs.

Each branch has a full environment : Agents launch isolated instances via Git worktrees and use Chrome DevTools Protocol for UI‑driven verification.

Mandatory automated validation : OpenAI validates via DevTools and observability tools; StrongDM uses external holdout scenarios, rejecting agents’ self‑reported confidence.

Agents reviewing agents : OpenAI pushes code review into a closed loop of agents; StrongDM eliminates human review entirely.

Continuous cleanup : OpenAI runs resident agents to scan for “AI waste” and opens small refactor PRs, a garbage‑collection‑style maintenance rather than periodic overhauls.

Why Now?

The breakthrough comes with GPT 5.2 Codex Extra High, a model capable of maintaining context over long spans, enabling agents to handle complex, multi‑step tasks end‑to‑end without supervision. A single prompt can drive a full bug‑to‑fix pipeline, running for six hours and delivering functional features while the team sleeps.

Author’s Current Workflow

At RetailBook, the author runs a large TypeScript monorepo (six apps, 20+ shared packages) with the following AI‑augmented steps:

Planning: Claude Opus 4.6 drafts the plan, capturing intent, edge cases, and integration points; GPT 5.3 Codex refines technical details.

Implementation: Codex writes code, tests, configs, and handles dependencies—no human code contributions.

UI Review: Opus scans generated Playwright screenshots to catch visual issues missed by Codex.

Code Review: Multiple GitHub‑Action‑driven agents review the PR, routing findings back to Codex for iterative fixes until all agents approve.

High‑risk changes still require human oversight, but most of the pipeline is fully automated.

Emerging Role: Senior Agent‑Native Product Engineer

The author describes a new hybrid role combining engineering, product management, and agent orchestration—designing systems, translating product intent into specifications, and building scaffolding that enables agents to produce high‑quality software.

Future Direction

The shared answer among the three teams is to enable agents to work reliably with minimal human intervention, shifting engineering discipline from code writing to environment design, feedback loops, and validation. This represents a second software‑industrial revolution: automating intellectual resources rather than just code.