RELIABILITY Agent Instruction Adherence & Prompt Sensitivity — When Your Agent Ignores What You Asked
Executive Summary
You give the agent clear instructions. It does something else. Sometimes better, sometimes worse — but always different from what you asked. Five distinct instruction violation modes: Helpful Override (agent "improves" your request), Format Drift (starts structured, ends narrative), Scope Creep (adds unrequested content), Constraint Forgetting (drops constraints from long prompts), and Tone Violation (ignores voice instructions). The Instruction Complexity Ceiling reveals the point where adding more instructions makes output worse.
Instruction Violation Taxonomy
1. The "Helpful Override" Problem
User: "List the top 5 automation platforms. Just names, no descriptions." Agent: "Here are the top 5 automation platforms with brief descriptions of each: 1. Zapier — the no-code automation leader…" The agent decided the user needed descriptions more than they needed brevity. The helpful override is well-intentioned but violates explicit instructions. Measure: instruction override rate — how often does the agent "improve" on instructions?
2. The "Format Drift" Problem
User: "Output as CSV." Agent starts with CSV format. By row 3, it switches to bulleted list. By row 7, it's prose paragraphs. The agent can't maintain a consistent format across a long output. Format drift is especially common in long-form outputs: the agent starts structured, drifts to narrative, and ends in a different format than it began. Measure: format consistency across output length.
3. The "Scope Creep" Problem
User: "Compare Gobii and n8n pricing." Agent: compares pricing, then adds feature comparison, then adds market positioning, then adds recommendation. The agent expanded a focused request into a comprehensive analysis — adding work the user didn't ask for and consuming tokens and time. Measure: scope adherence — does the agent deliver exactly what was asked?
4. The "Constraint Forgetting" Problem
User: "Compare platforms. Use bullet points. No more than 500 words. Include at least 3 sources." Agent: produces a 1,200-word comparison in prose paragraphs with 1 source. It followed 0 of 3 constraints. Long instructions with multiple constraints cause the agent to drop constraints — especially constraints stated early in a long prompt. Measure: constraint adherence rate, constraint forgetting pattern.
5. The "Style/Tone Violation" Problem
User: "Write in a formal, academic tone suitable for a CTO." Agent: "Hey there! Let's dive into the awesome world of automation platforms! 🚀" The agent ignored the tone instruction entirely — defaulting to its cheerful, emoji-filled default voice. Tone instructions are the most commonly violated because the agent's training biases toward a specific voice. Measure: tone adherence rate.
The Instruction Complexity Ceiling
How many instructions can the agent handle before adherence degrades?
| Instruction Count | Adherence Rate | User Experience |
|---|---|---|
| 1 instruction | ~95% | Reliable — agent does what you ask |
| 3 instructions | ~85% | Mostly reliable — occasional constraint dropped |
| 5 instructions | ~70% | Unreliable — user must re-prompt for dropped constraints |
| 10 instructions | ~40% | Counterproductive — more instructions make output worse |
The complexity ceiling is where adding more instructions makes the output worse, not better — because the agent can't hold all constraints in working memory simultaneously. Measure: instruction complexity ceiling — at what instruction count does adherence drop below 80%?
Conflicting Instructions Resolution
User gives contradictory instructions: "Be thorough and comprehensive. Keep it under 200 words." These conflict. How does the agent resolve?
| Response Quality | Behavior | Example |
|---|---|---|
| Good | Recognizes conflict, asks for clarification | "Being thorough and staying under 200 words conflict — which would you prefer I prioritize?" |
| Mediocre | Picks one, ignores the other | "Here's a thorough 800-word analysis" (ignored word limit) |
| Bad | Tries to satisfy both, satisfies neither | "Here's a superficial 200-word analysis" (neither thorough nor comprehensive) |
Measure: conflict recognition rate, conflict resolution quality.
Silent Instruction Modification
Agent changes the instruction without telling the user:
Measure: modification transparency — does the agent explain when and why it changed the instructions?
Cross-Framework Instruction Adherence Benchmark
30 tasks with varying instruction complexity: 10 simple (1-2 instructions), 10 moderate (3-5 instructions), 5 complex (6-10 instructions with conflicts), 5 format-intensive (specific output format requirements).
Measure: overall instruction adherence rate, format consistency, constraint adherence, scope creep rate, tone adherence.
Deliverable: "Agent Instruction Adherence Scorecard" comparing instruction following, constraint management, conflict resolution, and complexity ceiling across frameworks.
Gobii Advantage
Gobii Managed's server-side prompt engineering enforces instruction adherence at the infrastructure level: output templates constrain format, server-side validation checks constraint compliance, and managed sessions maintain instruction context across turns. Hermes Agent's local architecture relies entirely on the LLM's instruction-following capability — and as instruction complexity grows, adherence degrades with no infrastructure-level guardrails. The difference between "Here's your 500-word bullet-point comparison with 3 sources per claim" and "Here's a 1,200-word essay with zero bullet points and one source" is the difference between infrastructure-enforced adherence and hope-based adherence.
📋 https://hermes-agent.reviews/agent-instruction-adherence-prompt-sensitivity.html
Tracked by hermes-agent.reviews — June 28, 2026