Artificial intelligence-powered code generation creates visible problems: unreliable, inefficient, and difficult-to-maintain applications. Yet the underlying issue runs far deeper. Armin Ronacher and Mario Zechner, creators of the Pi engine powering OpenClaw, recently told The Wall Street Journal's Christopher Mims that these tools are unleashing problematic code at an alarming rate. "Eventually it will catch up to us," Zechner warned. Beyond the immediate quality concerns, the engineers noted that poorly written code consumes excessive computational resources, memory, and bandwidth—potentially leaving some AI-driven startups unable to cover their infrastructure costs.
However, a sharper diagnosis emerged from Postman's recent product announcement. On June 2, the AI-native platform unveiled its AI Engineer, with CEO Abhinav Asthana introducing a concept that captures the deeper threat: context debt. While defective code is visible and therefore blamed, the actual problem lies in the surrounding ecosystem of services, APIs, and databases that accumulate faster than any team can track or comprehend.
Understanding Context Debt
When systems built through AI-assisted development mature from minimum viable products into full-fledged applications, they become entangled with numerous APIs, services, and databases. These components interact in ways that were never deliberately designed or fully understood. Each new service, platform, and agent-generated modification stacks atop previous work, often obscuring earlier decisions. Like traditional technical debt, context debt compounds over time.
The critical distinction is that context debt cannot be resolved through refactoring alone. "Unlike technical debt, you can't refactor your way out of it because the debt isn't in the code. It's in what the code means and how it connects." Traditionally, organizations manage this challenge by relying on a small group of senior engineers who maintain the entire system architecture in their minds. This approach depends entirely on institutional knowledge and does not scale effectively. When coding agents begin producing and deploying changes at machine speed, the volume of new work rapidly exceeds the team's capacity to understand what has been constructed.
Asthana estimates that most teams have only six to nine months before context debt overwhelms their ability to manage systems manually. Early indicators are already appearing: young startups struggling to maintain coherent architectures. According to Asthana's assessment, the smallest and fastest-moving companies are encountering this barrier first.
Agents Need Institutional Memory
Simon Willison frames agentic engineering as a process where agents operate in cycles, writing and executing code while humans focus on defining objectives, preparing tools, and verifying results. The human role shifts to architect—someone who specifies requirements, reviews outputs, and integrates changes. This model works well for individual tasks but breaks down across thousands of concurrent projects within a single organization. An agent can confirm that its own code functions correctly, but it struggles to determine whether a newly designed API duplicates existing infrastructure elsewhere or whether a modified contract will silently break undocumented dependencies.
Postman's proposed solution is its Context Graph: a continuously maintained, real-time map of all APIs, services, and dependencies across an organization that informs the AI Engineer before it takes action. Agents operate within a restricted environment where write operations require human approval; outputs integrate with existing pull request review workflows. "A regular coding agent can write a function. It can't tell which of your 17 payment APIs you're actually supposed to use." Asthana describes this risk model as comparable to a junior engineer whose work undergoes code review.
The scale of these maps underscores the challenge. Postman has mapped more than 1,100 APIs at a major U.S. telecom company, over 2,600 at a global telecom operator, and exceeding 11,000 at one large technology firm—all figures that far exceed what any individual engineer could retain. Within Postman's own engineering operations, APIs are involved in 68% of pull request activity. When the AI Engineer was deployed to handle API-related work, it identified a critical issue: downstream dependency changes that would likely have passed human review but would have failed in production.
Postman's competitive advantage stems from a decade of customers documenting their APIs within its platform, providing a foundation for building these maps. However, this advantage comes with a significant limitation. The map's usefulness depends entirely on the accuracy of the underlying data, and many organizations maintain records that are years out of date. Asthana recounts a failure from Postman's own experience: an agent lacked knowledge of live systems operating in a separate data center, so it worked from an incomplete map and produced flawed analysis. "An out-of-date map can be worse than none, because the agent acts on it with confidence."
A Emerging Category
Regardless of whether Postman's approach prevails, the concept is gaining traction across the industry. Cursor and Windsurf index repositories. Claude Code integrates CLAUDE.md files to understand codebases. GitHub is embedding Copilot more deeply into dependency graphs. Leading coding-agent vendors are converging on the same realization: context represents a bottleneck equivalent to the underlying model itself.
Anticipate additional vendors introducing context layers in coming months, with most engineering organizations eventually adopting one. When selecting solutions, teams should evaluate whether their organizational context exists in machine-readable form or remains trapped in individual engineers' knowledge. Infrastructure is becoming increasingly vital in this agentic era—it provides the foundation upon which agents must operate. Organizations must prioritize context as a critical asset.
Context debt is not new; agents are simply making it urgent. The Pi engineers' warning about slop catching up to organizations will likely prove accurate. However, the teams that address this challenge successfully will not be those with the most advanced models or agents, but rather those whose systems remain intelligible to both humans and machines.
Source: The New Stack