Agent & Prompting Failures

Definition

The instruction is unclear, causing poor output.

Solution

Specify the goal, context, constraints, audience, and desired format.

Definition

The prompt contains too much information for the model to use effectively.

Solution

Remove noise and prioritize the most relevant context.

Definition

The input exceeds the model's maximum context length.

Solution

Use chunking, summarization, retrieval, or hierarchical processing.

Definition

Examples in the prompt confuse the model or teach the wrong pattern.

Solution

Use clear, consistent, relevant examples and test them across cases.

Definition

Important instructions are weakened by too much surrounding text.

Solution

Place critical rules clearly and use structured sections.

Definition

A user or retrieved document tries to override trusted instructions.

Solution

Use instruction hierarchy, content isolation, filtering, and output validation.

Definition

Hidden or private prompt instructions are exposed.

Solution

Avoid sensitive prompt content and filter outputs for prompt disclosure.

Definition

Small wording changes cause large output differences.

Solution

Test prompt variants and use robust templates.

Definition

The prompt works only for narrow examples and fails on real cases.

Solution

Evaluate prompts against diverse, realistic, and adversarial examples.

Definition

The prompt contains conflicting rules.

Solution

Simplify the prompt and prioritize instructions clearly.

Definition

The model cannot tell which instruction matters most.

Solution

Use hierarchy, numbered requirements, and examples.

Definition

The prompt works in demos but breaks with real users.

Solution

Use prompt evaluation sets and continuous prompt improvement.

Definition

The agent creates a poor, incomplete, or impossible plan.

Solution

Use planning constraints, examples, plan validation, and step limits.

Definition

The agent shifts away from the original objective.

Solution

Restate the goal and check progress after each step.

Definition

The agent keeps repeating steps without finishing the task.

Solution

Add loop limits, stopping rules, and escalation paths.

Definition

The agent keeps retrying failed actions without escalation.

Solution

Use retry limits, fallback strategies, and failure escalation.

Definition

The agent chooses or performs the wrong action.

Solution

Add action validation, safer tool APIs, and approval gates.

Definition

The agent fails to choose the best available tool.

Solution

Improve tool descriptions, routing logic, examples, and tool-use evaluation.

Definition

A tool call fails because of bad input, timeout, or system error.

Solution

Use input validation, retries, error handling, and timeout policies.

Definition

The agent uses a tool incorrectly.

Solution

Use strict schemas, tool-specific guardrails, and test cases.

Definition

The agent reads the tool result incorrectly.

Solution

Use structured tool outputs and verification prompts.

Definition

The agent tries to use a tool it is not allowed to use.

Solution

Apply least privilege and permission checks before tool calls.

Definition

Stored memory becomes inaccurate, outdated, or misleading.

Solution

Use memory validation, expiration, review, and user control.

Definition

Private or irrelevant memory appears in an answer.

Solution

Use memory access rules, privacy filters, and memory scoping.

Definition

Saved knowledge becomes less accurate over time.

Solution

Refresh, expire, or revalidate memory regularly.

Definition

The agent fails to retrieve useful memory.

Solution

Improve memory indexing, relevance scoring, and retrieval tests.

Definition

The agent does not check whether its answer or action is correct.

Solution

Add verifier steps before final output or external action.

Definition

The agent reviews its work but fails to catch mistakes.

Solution

Use external checks, automated tests, or human review.

Definition

The agent takes unsafe or unnecessary actions without supervision.

Solution

Add approval gates, scope limits, and risk-based controls.

Definition

The agent has too much freedom and too few limits.

Solution

Use permission boundaries, budgets, step limits, and human approval for risky actions.

Definition

The agent accesses or changes resources it should not.

Solution

Apply least-privilege access and enforce tool permissions.

Definition

The agent loses track of task state, progress, or decisions.

Solution

Use explicit state tracking, checkpoints, and structured task memory.

Definition

Multiple agents disagree, duplicate work, or block each other.

Solution

Define clear roles, shared state, coordination protocols, and conflict resolution.

Definition

One agent passes incomplete or incorrect context to another agent.

Solution

Use structured handoff summaries and required context fields.

Definition

The agent should ask for human help but continues alone.

Solution

Use uncertainty thresholds, risk rules, and escalation triggers.

Definition

One failed tool call breaks the entire workflow.

Solution

Design recovery paths, partial success handling, and fallback tools.