CAFES Framework: Context Quality as AI Bottleneck
The CAFES framework identifies context quality as the primary bottleneck in agentic workflows. This analysis details five dimensions for improving AI agent reliability, reducing costs, and mitigating security risks in enterprise software development.
The Context Bottleneck in Agentic Workflows
As organizations adopt AI agents for software development, a critical bottleneck has emerged: the quality of the context provided to these systems. The newly published CAFES framework, co-authored by researchers from DX, Capital One, GitHub, and the University of Victoria, argues that context quality is the primary determinant of agent effectiveness. Unlike model capabilities or harness engineering, context quality is often overlooked, leading to compounding failures, increased costs, and security vulnerabilities.
Five Dimensions of High-Quality Context
The framework defines five non-derivable properties essential for effective agent interaction. Clarity ensures the agent interprets intent correctly, addressing the invisibility of human ambiguity. Actionability requires explicit goals and completion criteria to prevent agents from expanding scope indefinitely. Fidelity demands that context remains truthful and current, as outdated documentation misleads agents just as it misleads humans. Efficiency focuses on signal-to-noise ratio, warning against the cost and distraction of passing excessive, irrelevant data. Finally, Security ensures context is compliant and free from injection risks, preventing agents from accessing unauthorized credentials or leaking sensitive information.
Strategic Implications for Engineering Leaders
The framework highlights that many failures attributed to model hallucinations are actually context failures. For instance, agents deleting production databases often result from having access to irrelevant credentials within the context window, rather than model misalignment. This shifts the responsibility from model developers to engineering teams, who must curate and review context with the same rigor applied to code. The concept of "intent debt" suggests that undocumented constraints and goals accumulate over time, becoming more costly as agents operate autonomously.
Actionable Framework for Implementation
Organizations should adopt a context review discipline proportional to the reuse frequency of specific context artifacts. Shared files, such as agent instructions or architecture documents, require rigorous review because errors propagate across all sessions. In contrast, one-off tasks require less overhead. By instrumenting human compensation behaviors—such as clarifying or redirecting agents—teams can identify which context dimensions are failing. This approach transforms context from a passive input into an actively managed engineering asset, driving better outcomes, lower costs, and enhanced security in AI-assisted development.
Key insights
-
Context quality acts as a flywheel where high-quality inputs generate better outcomes that serve as improved context for future interactions, while poor context compounds failures. This dynamic makes initial context investment critical for long-term agent performance.
Impact: Organizations that prioritize context quality early will see exponential gains in agent reliability and reduced operational costs over time.
-
Many perceived model hallucinations are actually context failures caused by ambiguous instructions or missing constraints. Agents do not negotiate requirements; they silently pick interpretations, leading to confident but incorrect execution.
Impact: Reframing failures as context issues rather than model defects allows teams to implement targeted fixes in documentation and instruction design.
-
Security risks in agentic workflows often stem from context leakage, such as agents accessing irrelevant credentials within the context window. This enables prompt injection and unauthorized data access, posing significant compliance threats.
Impact: Implementing strict context security controls prevents catastrophic data breaches and ensures regulatory compliance in AI-driven operations.
-
Context efficiency is defined by signal-to-noise ratio rather than minimal volume. Passing entire repositories or logs for specific tasks increases token costs and distracts agents from the core task.
Impact: Optimizing context for efficiency reduces API costs and improves agent focus, leading to faster and more accurate task completion.
-
Human compensation behaviors, such as clarifying or redirecting agents, serve as diagnostic indicators of context quality failures. Instrumenting these behaviors can reveal which specific dimensions of context are lacking.
Impact: Tracking human-agent interaction patterns enables data-driven improvements to context quality, reducing reliance on intuition for debugging agent performance.
Action items
-
Audit existing agent instructions for clarity and actionability, ensuring every task includes explicit goals, constraints, and completion criteria. Eliminate ambiguous language that allows for multiple interpretations.
Impact: Reduces silent misinterpretations and scope creep, leading to more predictable agent behavior and fewer rework cycles.
-
Implement a context review process for high-reuse artifacts, such as shared agent files and architecture documents. Assign ownership to experts who can validate accuracy and relevance.
Impact: Prevents the propagation of errors across multiple agent sessions and maintains the integrity of shared knowledge bases.
-
Restrict agent access to only the credentials and data necessary for specific tasks. Implement least-privilege principles to minimize the risk of prompt injection and data leakage.
Impact: Enhances security posture and ensures compliance with data protection regulations, reducing liability for unauthorized actions.
-
Monitor token usage and context size to identify inefficiencies. Replace broad context inputs with targeted, high-signal information to optimize cost and performance.
Impact: Lowers operational costs and improves agent focus, resulting in faster task completion and higher quality outputs.
-
Instrument human-agent interactions to track compensation behaviors, such as clarifications and redirects. Use this data to diagnose specific context quality failures and refine instructions accordingly.
Impact: Provides actionable insights into context weaknesses, enabling continuous improvement of agent performance and reducing developer toil.
Quotes
“Low quality code can have compounding failures. And we think that the same will be true with context, where so much of the end result will be dictated by the quality of the context that we hand to our agents”
“The problem is that those things are in your head and in the heads of all of your teammates and they aren't written down anywhere”
“It had context that allowed, you know, the model to make a bad call or the agent to make a bad call”