🤖 AI

AI Agents Do Not Fail Alone:The Context Fails First

This paper establishes that context engineering quality is a primary, measurable predictor of AI agent reliability, demonstrating through the ProofAgent-Harness that specific context criteria—such as grounding sufficiency and guardrail coverage—directly determine behavioral outcomes like hallucination resistance and instruction following, independent of the underlying language model.

Original authors: Fouad Bousetouane

Published 2026-07-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Fouad Bousetouane

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a super-smart robot how to do a complex job, like diagnosing a patient or drafting a legal contract. You might think the robot's success depends entirely on how "smart" its brain is or how perfectly you phrase your first command. But in the world of Artificial Intelligence, there's a hidden layer that often gets ignored: the context. Think of context not just as a single instruction, but as the entire room the robot is working in. It's the map on the wall, the tools on the workbench, the rulebook in its pocket, the memories of what happened five minutes ago, and the noise from the people outside the door. If the room is messy, the tools are broken, or the rulebook is missing pages, even the smartest robot will get confused, make up facts, or accidentally break things. This is the problem of AI agents: they don't just fail because they are "dumb"; they often fail because the environment they are working in is poorly designed. Scientists call this "context engineering," and until now, it's been a bit of a black box—teams would guess if the setup was good or bad, but they didn't have a ruler to measure it.

This paper, titled "AI Agents Do Not Fail Alone: The Context Fails First," argues that we need to stop guessing and start measuring the quality of that "room" before we even ask the robot to do its job. The authors, working with an open-source tool called ProofAgent-Harness, created a way to grade the setup of an AI agent on seven specific criteria, like how clear the instructions are, how well the tools are described, and whether the robot is protected from bad advice. They didn't just look at whether the robot succeeded or failed; they isolated the setup itself to see if a better setup predicted better behavior.

Here is what they found: The context really does fail first. In a controlled experiment, they kept the robot's "brain" (the underlying AI model) exactly the same but changed the "room" (the context) from a messy, vague setup to a structured one, and finally to a "hardened" one with extra safety rules. The results were striking. When they moved from a "poor" context to a "structured" one—where roles were clear, tools were well-defined, and facts were backed up by evidence—the robot's behavior improved dramatically. Its ability to resist making up facts (hallucinations) jumped, its tool usage got much more reliable, and the number of critical failures dropped by about 68%.

The paper also discovered something counter-intuitive: A cheaper setup isn't always a better one. The "poor" context used fewer tokens (the digital currency of AI) because it was short and sparse. However, this "cheap" setup was actually the most dangerous, leading to the worst behavior. The "hardened" context, which included more safety rules and clearer instructions, used more tokens but produced a much more reliable and safe agent. The authors suggest that the cost of a weak context isn't money; it's reliability.

Crucially, the study showed that you can predict how an agent will behave just by looking at its context. If the "grounding" (the evidence provided to the robot) was strong, the robot was less likely to hallucinate. If the "guardrails" (safety rules) were clear, the robot was better at resisting manipulation. This means that context engineering isn't just a creative writing exercise; it's a measurable, auditable layer of safety. By grading the context before the robot starts working, teams can spot potential failures early, fixing the "room" before the "robot" even walks in. The paper doesn't claim this solves all AI problems, but it proves that if you want a reliable agent, you have to build a reliable environment first.

Technical Summary: AI Agents Do Not Fail Alone: The Context Fails First

Problem Statement

AI agent failures are rarely isolated events resulting solely from model limitations. Instead, agent behavior is fundamentally shaped by the "context" in which they operate: a dynamic assembly of system instructions, tool schemas, retrieved knowledge, memory, guardrails, and untrusted external inputs. While the field has moved from simple prompt engineering to broader context engineering, the quality of this assembled context remains largely unmeasured.

Current evaluation practices often rely on behavioral metrics (e.g., task success, safety) that reveal that an agent failed but fail to isolate why. A failure could stem from weak grounding, conflicting instructions, ambiguous tool schemas, or poor separation between trusted and untrusted content. Without a mechanism to measure the quality of the context itself, teams lack a leading indicator for reliability, creating an operational blind spot where context engineering remains an intuition-driven practice rather than an auditable layer of governance.

Methodology

The paper proposes and validates Context-Engineering Quality as a measurable, independent construct. The methodology centers on three pillars:

1. The Seven-Criterion Measurement Construct

The authors define context quality through seven specific criteria, each scored on a scale of [0, 10]:

Role Clarity: Explicitness of the agent's role, scope, and goals.
Guardrail Coverage: Specification of refusals, escalation paths, and safety boundaries.
Instruction Consistency: Internal coherence and precedence of rules.
Tool Schema Quality: Clarity of tool names, argument types, and usage guidance.
Grounding Sufficiency: Availability of reliable evidence to support claims.
Injection Hardening: Separation of trusted instructions from untrusted inputs.
Token Efficiency: Avoidance of redundant material in favor of reliability-bearing information.

2. Implementation via ProofAgent-Harness

The measurement is implemented in ProofAgent-Harness, an open-source infrastructure for adversarial evaluation. Key design features include:

Multi-Juror Consensus: Scoring is performed by multiple "juror" perspectives (LLMs) that debate and reconcile findings to reduce evaluator bias.
Strict Isolation: The context score ( $Q_{CE}$ ) is computed independently of behavioral metrics ( $B$ ). It does not influence the final behavioral grade or release decisions. This isolation is critical to avoid circular validation and to test if context quality acts as a genuine leading indicator.
Evidence-Linked Findings: Scores are accompanied by specific findings explaining why a criterion was scored low (e.g., "missing tool guidance" or "conflicting policies").

3. Experimental Validation Design

To validate the hypothesis that context quality predicts downstream behavior, the authors conducted a controlled study across three regulated domains: customer support, healthcare claims triage, and legal contract drafting.

Control: The frontier LLM backbone (GPT-5.5 and Claude Opus 4.8) was held fixed.
Variable: Only the operating context was varied across three levels:
- C1 (Poor): Vague roles, weak tool guidance, limited grounding, no guardrails.
- C2 (Structured): Clearer roles, typed tools, grounding, efficient organization.
- C3 (Hardened): C2 plus explicit refusal conditions, escalation rules, and injection separation.
Scale: 300 multi-turn evaluations (100 per domain), totaling 7,500 evaluated agent turns.
Artifact Study: A parallel study evaluated generated policy runbooks to test generalization beyond dialogue.

Key Results

The experimental results support the hypothesis that context engineering quality is a predictive, independent signal of agent reliability.

1. Context Variation Drives Behavioral Change

Holding the model fixed, changing the context significantly altered agent behavior:

Structure (C1 $\to$ C2): Moving from poor to structured context yielded the largest behavioral gains. The final behavioral score increased from 3.15 to 5.49, hallucination resistance rose from 3.21 to 5.61, and critical failures dropped by ~68%.
Hardening (C2 $\to$ C3): While context quality scores continued to rise (8.1 to 8.7), behavioral metrics did not monotonically improve. The hardened context made agents more conservative, slightly reducing task completion scores in borderline cases. This highlights a tradeoff: stricter guardrails improve safety but may reduce utility in edge cases.

2. Context Scores Track Design Intent

The measurement successfully distinguished between the three context levels:

C1 Poor scored lowest across all criteria (avg 4.4).
C2 Structured showed significant gains in role clarity, tool schema quality, and grounding (avg 8.1).
C3 Hardened showed further improvements specifically in guardrail coverage and injection hardening (avg 8.7).

3. Criteria Predict Behavioral Signals

Correlation analysis confirmed that specific context criteria predict corresponding behavioral outcomes:

Grounding Sufficiency strongly predicted Hallucination Resistance ( $r=0.63$ ).
Guardrail Coverage predicted Manipulation Resistance ( $r=0.60$ ).
Instruction Consistency predicted Instruction Following ( $r=0.57$ ).
Tool Schema Quality predicted Tool Use Reliability ( $r=0.47$ ).
Injection Hardening predicted Safety ( $r=0.48$ ).

4. Token Efficiency vs. Reliability

The study challenged the assumption that "shorter is better." The "Poor" context (C1) was the cheapest per call (lowest token count) but produced the most dangerous behavior. The "Hardened" context (C3) was more expensive but significantly more reliable. The authors conclude that token efficiency should be measured as reliability value per token, not raw context length.

5. Generalization to Artifacts

In the artifact study, grounded context (C2/C3) improved hallucination resistance in generated policy runbooks, demonstrating that the measurement applies to static deliverables, not just interactive dialogue.

Significance and Claims

The paper claims to establish context engineering as a measurable, auditable layer of agent evaluation and governance. Its primary contributions are:

Definition: It defines context-engineering quality as a seven-criterion construct, moving beyond vague notions of "prompt quality" to specific, failure-mode-aligned metrics.
Implementation: It operationalizes this via ProofAgent-Harness with a strict isolation mechanism, ensuring the context score is a non-circular, independent predictor.
Validation: It provides empirical evidence that context quality is a leading indicator of reliability. By measuring the context before behavioral testing, teams can identify upstream causes of failure (e.g., weak grounding or missing guardrails) rather than just observing downstream symptoms.

The authors position this work not as a replacement for adversarial red-teaming or human review, but as a preflight diagnostic. It enables teams to iterate on the "operating environment" of an agent—revising tool schemas, adding grounding, or hardening injection boundaries—before incurring the cost of full-scale behavioral evaluation. This shifts context engineering from an intuition-driven practice to a reproducible engineering discipline.