Agencies are starting to treat AI coding tools less like software and more like junior developers: you describe the thing, steer when it drifts, and something functional eventually appears. The article frames this as “vibe coding”—tools like Replit, Anthropic’s Claude Code, and OpenAI’s Codex letting non-engineers build internal tools and campaign assets without waiting on dedicated dev resources.
At Anomaly, global chief AI officer Christopher Neff used Replit to build an awards “submission engine” that analyzes creative and recommends categories across shows like Cannes Lions, the Clios, and AICP, then helps shape entries to match submission rules. The workflow is stitched together from multiple systems: Perplexity’s Computer agent compiled roughly five years of award data and judges’ commentary into a ~1,500-page research base; Claude turned that into a structured dataset; and Replit handled the build. Neff’s role was supervisory—catching issues like sequential (instead of parallel) analysis and API misfires (including with GitHub) and pushing the system toward something usable. It’s less “press button, receive app,” more “manage the robot intern until it stops eating staples.”
Code and Theory applied a similar approach closer to production. Two designers with limited engineering experience used Claude Code to build an effects platform for an Adobe campaign in a single day, generating assets with a computer-vision-inspired look—blob tracking, ASCII-style references, heavy pixelation. The process was iterative: upload references, test outputs, adjust via prompts. When errors appeared, they described them back to Claude, which handled debugging. (“We used Claude to troubleshoot Claude.”)
Elsewhere, indie agency Gus used Codex (with help from an external AI specialist) to build a Slack bot, Sug, that runs daily check-ins and sends a morning report to leadership. The bot asks employees to report on project status via quick inputs, flagging potential issues for follow-up. According to Gus, the system can surface problems earlier and may save “two or three days” of misalignment or team stress.
The shift here isn’t that agencies suddenly became engineering organizations. It’s that more problems start to look small enough to try building—and fast enough to attempt without a full scoping exercise. Whether the output behaves like production software or a well-managed prototype seems, for now, to depend on how closely someone is willing to supervise the “developer.”
Read more at AdAge.