The "Context Bloat" Tax:
Are Your AI Coding Agents Are Costing More and Delivering Less?

25 February 2026

by Srikumar Sundareshwar, CTO

​

As practitioners of Enterprise AI adoption, we are moving past the "AI hype" phase and into the Optimisation Phase. We are deploying sophisticated agents like Claude 4.6 Sonnet and GPT-5.2 into our production pipelines. To guide them, we’ve adopted the current industry "best practice": providing repository-level context files like AGENTS.md or CLAUDE.md. We thought we were providing a "North Star."

​

New research suggests we are actually giving them an expensive anchor. A new paper from researchers at ETH Zurich (arXiv:2602.11988) reveals that these context files often act as Distraction Anchors that cannibalise your ROI.

 

The Finding: Efficiency ≠ Effectiveness

It is vital to understand that the problem isn't the accuracy of the context file. The agents do follow the instructions. The problem is the economic trade-off. The study found:

 

• Negative ROI on Automation: Using LLMs to auto-generate repo summaries led to a 3% decrease in task success.

• The 20% "Bloat Tax": Including these files increased inference costs by more than 20% across the board.

• The Compliance Trap: Agents spend significantly more tokens on "file traversal" and "thorough testing" to satisfy the documentation's requirements, which rarely leads to a faster or better fix.

 

The Industry Disconnect: Strategy vs. Data

These findings create a direct challenge to the official playbooks from major providers:

​

• OpenAI: Traditionally advocates for AGENTS.md to "accelerate AI onboarding". The data shows this "onboarding" often triggers "correctness loops" where agents prioritise process over the P0 bug.

• Anthropic: Recommends a CLAUDE.md to define the project’s "WHAT, WHY, and HOW." This research proves that even these well-intentioned overviews can cause agents to "over-reason," driving up costs without moving the needle on success.

​

ROI Must Be Front-and-Center

As leaders, it's not sufficient to only ask "Is the agent right?" (as important as the question is & its vitally important) and it's also vital that we start asking "Was the fix cost-effective?"

​

If an agent takes 20% more tokens to reach the same result because it was busy "verifying its onboarding," the ROI of that automation has just evaporated. In a multi-agent enterprise ecosystem, "Context Bloat" isn't just a technical quirk—it's OpEx leakage.

​

Staving Off "Context Rot"

Beyond the bill, we are fighting Context Rot. Every token used by a bloated AGENTS.md file is a token stolen from the agent's active reasoning. As the context window fills with "noise," agents enter the "Dumb Zone"—a state where retrieval accuracy drops and the model starts "forgetting" the core problem to focus on minor style rules.

​

The EA Playbook: Architecting for Lean Context

To protect your ROI, you must treat your context files as performance-sensitive artifacts rather than standard documentation. We should apply the same rigorous "pruning" to our agent instructions that we apply to our production code. Here is how to re-architect your context strategy for maximum efficiency:

 

1. Enforce the "Discoverability Rule": If an agent can find information using existing tools (like ls, tree, or grep), keep it out of the file. There is no reason to pay the "Bloat Tax" to tell an agent about a directory structure it can fetch "just-in-time" for a fraction of the cost.

2. Document Anomalies, Not Standards: Skip the generic architectural overviews. Only include "weird" project quirks that the code cannot explain (e.g., "The /legacy folder is a trap; use /v2 for all new features").

3. Prioritize Action over Philosophy: Agents respond better to commands than values. Replace "We value modularity" with "Run ./test.sh --fast to verify changes."

​

The Bottom Line: Successful enterprise AI adoption isn't just about how much context you can give an agent—it's about the Information Density of your instructions. Are you optimising for a "successful fix," or are you accidentally paying a 20% premium for "Cognitive Noise"?

​

Original posting LinkedIn - Read Here