AI agents that build their own tools are already here, and Premiere Pro is the proof

Andrew Ambrosino offered a detail that deserves more attention than it has received. OpenAI’s Codex, he reported, did not merely write code inside an editor. It built its own bridge into a third-party application, installed that bridge, and then used it to issue commands back to the host software. The specific demonstration involved Adobe Premiere Pro: Codex constructed a Premiere Pro extension, installed it into the running application, and used the extension to manipulate markers inside the timeline.

That sequence matters because it describes something qualitatively different from code completion or script generation. A system that identifies a capability gap, writes the tooling required to close it, and then operates through that tooling is not just faster than a human developer. It is doing something a human developer would have needed explicit direction and several working sessions to accomplish. The gap Codex diagnosed and closed, in this case, was the absence of a programmatic channel between itself and Premiere Pro’s internal state.

The architecture Ambrosino described is not speculative. Multiple independent developers have published working implementations of precisely this pattern on GitHub. Repositories from developers including antipaster, isaiahdupree, and others provide extension-and-server bridges that allow AI agents to control Adobe Premiere Pro from outside the application, including the manipulation of markers, through an installed extension acting as a communication layer. The antipaster repository, available publicly, frames the project explicitly as an MCP server for Premiere Pro control. None of these projects appear to be affiliated with OpenAI or Adobe. They are independent corroboration that the technical approach Codex used is real, reproducible, and already being pursued by the broader developer community without coordination.

Codex then did was built itself an extension that could be installed into Premiere Pro that it could then talk to and say, "Hey, Premiere Pro extension, can you please change this marker inside of the Premiere Pro app." Andrew Ambrosino

What the GitHub activity establishes is that Ambrosino’s account is not an isolated anecdote about a capability that exists only inside a research environment. The extension-bridge pattern for AI-to-Premiere-Pro control has been implemented, shared, and is presumably being iterated on by developers who arrived at the same architecture independently. That is a meaningful signal about where agentic AI tooling is headed, not because any single repository is significant on its own, but because the simultaneous appearance of the same approach across unconnected projects suggests the pattern is both obvious enough to rediscover and functional enough to publish.

The Premiere Pro case is worth examining as a template. Creative professional applications like Premiere Pro are not designed to be controlled by external agents. They expose limited scripting interfaces, and their internal state is largely opaque to software running outside the process. The extension-as-bridge approach solves that problem by putting a communication endpoint inside the application itself, one that the AI agent can then address as if it were calling an API. Codex, in Ambrosino’s account, derived this solution without being handed it. That derivation is the part the benchmark scores and capability evaluations tend not to capture cleanly.

The broader implication is that the relevant unit of AI capability may be shifting. The question is no longer only what a model can do with tools it has been given. It is whether a model can identify which tools it lacks, build those tools, and integrate them into a working pipeline. If the Premiere Pro demonstration is representative rather than exceptional, the set of tasks that require a human to first design the scaffolding before AI can assist is shrinking faster than most workflow assessments have assumed. The GitHub repositories suggest that developers, at least, are already designing around that expectation.