The beanbag chairs were orange, the carpets were shag, and Adele Goldberg and Alan Kay were talking about a “personal dynamic medium” that functioned less like a tool and more like an organism.
In the early 1970s at Xerox PARC, the vision for computing wasn’t about text files or directory trees. It was about the System Image. In environments like Smalltalk-80, you didn’t “write code” in the modern, disconnected sense. You inhabited a live environment where the code, the data, and the execution state were a single, indivisible entity. You didn’t save a file; you saved the image—a snapshot of a living system.
Then, for fifty years, we lost the thread. We traded the living image for the Unix philosophy: “Everything is a file.” We moved into an era of stateless, “dead” text—passive archives of ASCII characters that only come to life when a compiler breathes temporary soul into them.
But as we enter the era of agentic software engineering, the ghost of the 1970s is returning. The repository is no longer a filing cabinet; it is becoming a stateful System Image once again.
The Stateless Dark Age
For years, the git repository has been treated as a passive storage unit. A developer pulls “dead” code out of the repo, “animates” it in their own mind to understand the logic, makes a change, and puts the dead text back.
Early AI coding assistants followed this same stateless pattern. They were essentially high-end autocomplete: they looked at the file you were currently editing, offered a suggestion, and then promptly forgot you existed. There was no “long-term memory,” no awareness of the project’s broader intent, and no persistence of state.
The Agentic Shift: Repo as Memory
The shift to agentic workflows—where AI agents don’t just suggest lines but actually “inhabit” the codebase—changes the fundamental architecture of the repository.
When an agent is tasked with evolving a system, it needs more than just access to the source code. It needs to understand the “why” behind the architecture and the “how” of the reasoning. To an agent, a repository isn’t a collection of files; it is its environment.
We are seeing the emergence of the Repo as Image paradigm. In this model, the repository functions as the agent’s Long-Term Memory (LTM). It stores not just the application logic, but the cognitive architecture of the project itself.
The New “Constitution”: AGENTS.md and Agent Skills
In the Smalltalk days, the “Image” held everything. In the modern agentic repo, we are beginning to formalize this via two key components that sit right alongside our source code:
- AGENTS.md (The Project Constitution): Instead of leaving an agent to guess the rules, we are now checking in a “manifesto” for the project. This file serves as the agent’s governing logic. It defines the architectural invariants, the stylistic soul of the project, and the specific personas allowed to operate within it. This is the modern equivalent of the Smalltalk environment’s self-documenting nature.
- Agent Skills (The Capabilities): If the constitution is the “mind,” Agent Skills are the “limbs.” These are the specific, tool-based capabilities checked into the environment—from specialized search heuristics to automated refactoring patterns. When an agent enters a repo, it “inherits” these skills, effectively becoming a specialist tailored specifically for that system image.
The Return of Bespoke Software
The beauty of the 1970s vision was that software felt personal. Because the image was live, you could mold it to your exact needs in real-time.
We are returning to that “bespoke” era. When the marginal cost of code evolution drops toward zero thanks to agents who “live” in the repo, the need for one-size-fits-all, packaged software diminishes. We can return to the idea of software that evolves continuously alongside the user.
The repository has become the “save state” of a living digital organism. We are no longer just “text editors”; we are system curators. We are finally building the dynamic medium that the pioneers at PARC dreamed of. The future of computing isn’t frozen artifacts and better autocomplete; it’s malleable, living systems. The Image is back.
We’re going back to the seventies, and this time, we’re staying.