Detailed Analysis
A recent case study circulating from OpenAI describes three engineers shipping an internal product—roughly 1,500 pull requests resulting in a codebase exceeding one million lines—almost entirely through AI agents, with individual Codex sessions running six or more hours at a stretch. The claim that this took "a tenth of the time" of manual development is less noteworthy than the operational problem the team encountered: as their agent runs grew longer, a single comprehensive instruction file meant to guide the project began to work against them. Rules written at the outset of the project accumulated into what the team called "a graveyard of stale rules," crowding out the actual task at hand and confusing the agent with directives that no longer matched the current state of the codebase. Their fix was architectural rather than incremental—replacing the monolithic instruction manual with a lightweight, continuously updated map pointing to active execution plans, decision logs, design documents, and quality assessments for each part of the system.
This pattern is not confined to OpenAI's tooling. Anthropic has converged on a strikingly similar solution in its own long-running agent work, particularly in scientific computing applications built on Claude Code. There, a "progress file" functions as portable memory between sessions, recording completed work, known limitations, and—critically—failed approaches along with the reasons they failed, so that a fresh Claude session doesn't repeat the same dead ends. Anthropic's long-running coding harness uses the same underlying structure: progress files, structured handoffs, and version history that let new sessions pick up where previous ones left off. The convergence of two competing labs on nearly identical mechanisms suggests this isn't a quirk of one company's tooling but a structural requirement that emerges once agents are asked to sustain work across many hours and multiple sessions rather than answer a single prompt.
The underlying issue is that context windows, however large, are not the real constraint on long-running agent work—context curation is. A giant, frozen instruction set written before a project begins can only encode what its authors knew at that moment. As agents execute and produce evidence—discovering that a research thesis doesn't hold, or that an early implementation choice creates a downstream dependency problem—the original brief becomes stale, sometimes actively misleading. The described method, "progressive context shaping," argues for treating the current state of a project as a living, prioritized document that supersedes the history of everything that came before it, rather than layering new instructions on top of old ones and hoping the agent sorts out which take precedence. This matters because it reframes what capability actually means for agents: it's not just about the model's raw reasoning or the size of its context window, but about the surrounding memory architecture that lets human judgment stay current as a task runs far longer than any person could supervise moment-to-moment.
The stakes attached to this shift are asymmetric in a way that makes the practice more than a workflow nicety. A well-directed long-running agent can carry a good decision much further than a chat-based interaction ever could—but it can just as easily carry a subtly wrong decision through hours of research, code, and drafts before the error becomes visible, at which point substantial work may need to be discarded. This is the same tension driving broader industry investment in agent memory, checkpointing, and handoff protocols: as models like Claude and GPT-based systems are increasingly deployed for multi-hour, multi-session autonomous work rather than single-turn assistance, the bottleneck shifts from model intelligence to the surrounding scaffolding that keeps that intelligence pointed at an accurate, current picture of the task. The fact that Anthropic and OpenAI have independently arrived at nearly the same solution—short, living state documents replacing exhaustive upfront specifications—signals that this scaffolding is becoming a recognized discipline in its own right, likely to be formalized into standard tooling as agentic coding and research systems become mainstream rather than experimental.
Read original article →