Why Your AI Agent Ignores Your Instructions (and How Custom Instructions and a Knowledge Base Fix It)

Caywork Platform

Author at Caywork

You set up an AI agent, gave it clear rules, and then watched it do something you specifically told it not to do. Maybe it answered a customer in the wrong tone, made up a price, or skipped a step you had written in capital letters. It is a frustrating moment, and it often leads people to decide that the technology simply does not work. In most cases, though, the problem sits in how the agent was set up, not in the AI itself. This guide explains, in plain language, why agents drift away from their instructions and how custom instructions, an AI agent knowledge base, and a few simple safety limits keep them on track. No technical background is needed to follow along.

Why AI Agents Seem to Ignore What You Tell Them

Before fixing anything, it helps to understand what is really happening when an agent goes off script. An AI agent does not read your rules the way a person reads a policy and then decides to break it. It works with whatever information is in front of it at the moment it answers, and that information is not always what you expect.

It Is Usually a Setup Problem, Not a Stubborn AI

The problem is real and growing. The Guardian reported in March 2026 on a study funded by the UK's AI Security Institute that documented nearly 700 real-world cases of AI chatbots and agents ignoring instructions or working around safeguards, with a fivefold rise between October 2025 and March 2026. Headlines like this can make agents sound unpredictable by nature. For everyday business use, the usual cause is far less dramatic: unclear rules, messy reference material, and agents given more freedom than their job requires.

The Three Things Behind Every Answer: Instructions, Knowledge, and the Request

Every time an agent responds, it works from three ingredients. The first is its instructions, which describe its role, tone, and rules. The second is its knowledge, meaning the documents and facts it can look up about your business. The third is the request itself, the question or task a person just gave it. When an agent misbehaves, one of these three is almost always unclear, missing, or in conflict with the others.

What This Guide Covers

The sections below start with the most common reasons agents go off script. They then show how to write instructions an agent will actually follow and how to build a knowledge base that supports those instructions instead of fighting them. The final part covers guardrails and blast radius, two ideas that decide how much damage a mistake can do, followed by a short checklist you can use whenever an agent misbehaves.

The Most Common Reasons an AI Agent Goes Off Script

Agents rarely ignore instructions at random. The same handful of causes shows up again and again, whether the agent answers customer questions, drafts content, or updates records. Knowing these patterns makes it much easier to find the real problem instead of rewriting everything and hoping for the best.

Instructions That Are Vague or Contradict Each Other

An instruction like “be helpful and professional” sounds clear to a person but gives an agent very little to work with. Contradictions are even harder: “keep answers short” in one place and “always explain every option in detail” in another forces the agent to choose, and it may not choose the one you meant. Gartner has pointed out that many of today's AI models still struggle to follow nuanced instructions over time, which is why every rule should be specific, easy to check, and consistent with the rest.

Too Many Jobs Packed Into One Agent

It is tempting to build one agent that answers support questions, writes social posts, and checks invoices. Every extra job adds more rules, and those rules start competing for the agent's attention. Anthropic, the company behind the Claude models, advises keeping the information an agent works with as small and focused as possible, because more is not always better. An agent with one clear job is easier to instruct, easier to test, and much less likely to mix up its rules.

Long Conversations Where Early Instructions Fade

Agents can lose track of the bigger picture as a conversation grows. Microsoft Research and Salesforce tested 15 leading AI models across more than 200,000 simulated conversations and found that performance dropped by 39% on average when a task was spread over several messages instead of given all at once. Once a model takes a wrong turn early, it tends to stay on that path. Chroma reached a similar conclusion from another angle: all 18 models it tested became less reliable as the amount of text they had to handle grew, even on simple tasks.

Missing or Outdated Information in the Knowledge Base

If an agent cannot find an answer in its knowledge, it may guess. A price list from last year, a product page that was never added, or a policy that changed last month all lead to confident but wrong answers. From the outside, this looks like the agent ignoring the rule “only use official information,” when in reality the official information simply was not there.

Documents That Quietly Overrule Your Instructions

Sometimes the knowledge base itself contains instructions. An old training manual might say “always offer a refund on request,” while the agent's instructions say “send refund requests to a manager.” The agent reads both and may follow the document. OWASP, the nonprofit behind widely used security guidelines, ranks prompt injection, where text the AI reads changes how it behaves, as the top risk for AI applications. Usually that means someone hiding instructions in a document, email, or web page on purpose, but an outdated manual can cause the same effect by accident.

Custom Instructions: Your Agent's Job Description

Custom instructions, sometimes called a system prompt, are the standing rules an agent carries into every conversation. Think of them as the job description you would hand a new employee on their first day. A good job description does not try to cover every possible situation; it explains the role, the priorities, and the lines that must never be crossed.

What Custom Instructions Are, in Plain Terms

Instructions tell the agent who it is, who it is helping, what a good result looks like, and what it should never do. They apply every time, whatever the user asks. If the knowledge base is the agent's reference library, the instructions are its manager's briefing: short, clear, and focused on how to behave rather than on facts.

What Belongs in Instructions and What Does Not

Instructions are the right place for the agent's role, tone of voice, step-by-step process, and firm rules such as “never quote a price that is not in the price list.” They are the wrong place for long product details, policy documents, or anything that changes often, because those belong in the knowledge base. Keeping the two apart means you can update a price list without touching the agent's behavior, and adjust its tone without editing a single document.

Before and After: Rewriting Instructions an Agent Will Follow

Small wording changes make a big difference. The examples below show how vague instructions can be rewritten into ones an agent can follow consistently. Each improved version tells the agent what to do, how much to say, and what to do when it is unsure.

Before (Vague)	After (Clear)
Be friendly and helpful.	Reply in a warm, plain tone in no more than four sentences. Use the customer's name if it is given.
Answer customer questions.	Answer questions about orders, delivery, and returns using only the documents provided. For anything else, say you will pass it to the team.
Don't make things up.	If the answer is not in the documents, say you do not have that information and offer to connect the customer with a person.
Handle refunds.	Never approve a refund yourself. Collect the order number and the reason, then send the request to a manager for approval.

The AI Agent Knowledge Base: What Your Agent Knows About Your Business

An AI agent knowledge base is the collection of documents and facts an agent can look up while it works: price lists, product details, policies, FAQs, and process guides. Without it, the agent falls back on general knowledge, which knows nothing about your business. With a good one, it can answer as if it had read your company handbook.

How a Knowledge Base Differs From Instructions

Instructions shape how the agent behaves, while the knowledge base shapes what it knows. A simple test helps: if a piece of text tells the agent to do something, it belongs in the instructions; if it describes a fact the agent may need, it belongs in the knowledge base. Mixing the two is one of the most common reasons agents seem to ignore their rules, because a rule buried inside a long document is easy to overlook or contradict.

Choosing the Right Documents to Add

More documents do not automatically make a smarter agent. Choose material that directly supports the agent's job, and leave out drafts, duplicates, and anything only loosely related. Clear, text-based documents with descriptive titles are easier for an agent to search than scanned images or complex layouts, and a short, accurate FAQ often helps more than a hundred-page manual.

Keeping Your Knowledge Base Clean, Current, and Consistent

A knowledge base needs regular care, just like a shared drive. Set a simple routine to remove outdated files, replace old price lists, and check that no two documents give different answers to the same question. Whenever a policy changes, update the knowledge base the same day, so the agent never has to choose between the old version and the new one.

Guardrails: Clear Limits on What an Agent Can Do

Guardrails are the limits that keep an agent inside safe boundaries, even when something goes wrong. They matter because instructions are requests, not locks. A well-written instruction makes good behavior likely, while a guardrail makes certain bad outcomes impossible.

Why Written Instructions Alone Are Not Enough

In 2026, six national cybersecurity agencies from the United States, the United Kingdom, Canada, Australia, and New Zealand jointly published guidance on adopting AI agents carefully, putting strict limits on access at the center. Kiteworks, which analyzed the guidance, sums up the lesson in one line: “System prompts are instructions, not controls.” In everyday terms, telling an agent “never delete customer records” is useful, but if the agent has no permission to delete records in the first place, it simply cannot, however confused it gets.

Rules Inside the Agent vs Limits Around It

Rules inside the agent are the instructions it reads. Limits around the agent are settings it cannot change: which tools it can use, which data it can see, and which actions it can take. Picture a new employee who has been told not to enter the finance office and has also not been given a key card for it. The instruction sets the expectation, and the missing key card makes sure it holds, which is why the strongest setups use both.

When a Person Should Approve Before the Agent Acts

Some actions are too important to leave to an agent alone, such as sending money, issuing refunds, deleting data, or emailing a large customer list. For these, a simple approval step works well: the agent prepares the action, and a person confirms it. FINRA, the US financial industry regulator, recommends this kind of human checkpoint in its 2026 oversight report before agents that can act or carry out transactions go ahead. The same idea fits almost any business: let the agent do the groundwork and keep the final click with a person.

Blast Radius: How Much Can Go Wrong When an Agent Makes a Mistake

Blast radius is a simple way to ask one question: if this agent makes a mistake, how far can the damage spread? An agent that can only read a product FAQ has a tiny blast radius. An agent that can read every customer record, send emails, and change orders has a very large one.

What Blast Radius Means in Everyday Terms

Picture two interns. One can only open the brochure folder; the other has keys to every office and the company credit card. Both might make the same small mistake, but only one of those mistakes can turn into a serious problem. AvePoint describes the blast radius of a compromised agent as reaching every system and dataset it was allowed to touch, which is why access matters as much as instructions.

Give Each Agent Only the Access Its Job Needs

The fix is to give each agent the smallest set of permissions it needs for its job, and nothing more. This matters because problems are already common: AvePoint found in its State of AI 2026 report that 88.4% of organizations experienced at least one AI agent-related security breach, and Gravitee reports that 88% of organizations confirmed or suspected an AI agent security or data privacy incident in the past year. IBM found that among organizations reporting a breach involving their AI, 97% lacked proper AI access controls. Limiting access does not prevent every mistake, but it keeps small mistakes small.

Start Small, Then Expand What the Agent Can Touch

A safe rollout grows step by step. Start with an agent that can only read information and draft suggestions, watch how it performs on real work, and only then let it take actions such as sending messages or updating records. QueryPie suggests setting limits in three ways, covering what the agent can access, for how long, and how much it can do, with the example of a sales agent that can read only sales customer data, until the end of the month, for up to 50 operations a day. Make sure you can pause an agent quickly too: Kiteworks found in its 2026 Forecast that 60% of organizations cannot quickly shut down a misbehaving agent.

A Simple Checklist When Your Agent Ignores Instructions

When an agent ignores its instructions, a calm, step-by-step check usually finds the cause faster than a full rebuild. The three steps below cover the most likely problems in the order you should check them. In most cases, the cause turns up in the first step.

Review the Instructions First

Read the instructions as if you were the agent seeing them for the first time. Look for vague words, missing steps, and rules that pull in different directions. These questions help you spot the gaps quickly.

Is the agent's job described in one or two clear sentences?
Does every rule say what to do, not only what to avoid?
Do any two rules contradict each other?
Is there a clear rule for what to do when the agent is unsure?
Are there rules for jobs this agent should not be doing at all?

Check What the Agent Can Actually Find

If the instructions look fine, the problem often lies in the knowledge base or in the length of the conversation. Check the information the agent relied on, not just the answer it gave. These questions point you to the most likely gaps.

Is the information the agent needed actually in the knowledge base?
Is it up to date, and is there only one version of it?
Does any document contain rules that conflict with the agent's instructions?
Has the conversation run long? Starting a fresh one is often the quickest fix, as the Microsoft Research and Salesforce study suggests.

Test With Real Examples Before Going Live

Before an agent goes live, and after every change, test it with real questions from your customers or team, including the awkward ones. Keep a short list of tricky cases, such as a refund request, a question the knowledge base cannot answer, and a request outside the agent's job, and run them every time you edit the instructions. If the agent handles all of them well, you can trust it with more; if not, you know exactly where to look.

Building Agents That Follow Instructions on the Caywork Platform

Everything in this guide applies to any AI agent, and it also reflects how Caywork approaches agents on its platform. Caywork is a marketplace and builder for AI agents, designed for businesses that want useful automation without a technical team. The aim is to make clear instructions, focused knowledge, and sensible limits part of how an agent is set up from the start.

Ready-Made Caywork AI Agents With Instructions Already Tuned

The quickest way to avoid instruction problems is to start with an agent that has already been built and tested for one specific job. Caywork's marketplace offers ready-made AI agents created around clear tasks, so the instructions, steps, and tools are already in place. You can try one on your own work and see how a focused agent behaves before deciding to build anything yourself.

Adding Your Own Instructions and Business Knowledge

When you need an agent that reflects your own business, Caywork's visual builder lets you set one up without writing code. You describe the agent's job and rules, add the business information it should rely on, and choose the tools it can use, which keeps its access limited to what the task needs. Following the approach in this guide, rules go in the instructions and facts go in the reference material, so each can be updated on its own.

Testing a Caywork Agent Before You Rely on It

New Caywork accounts receive 50 free credits with no card required, which is enough to run an agent on real examples before trusting it with daily work. Credits are one-time purchases with no subscription, so you can test, adjust the instructions, and test again while seeing exactly what each run costs. That makes the checklist above easy to follow in practice.

Try a ready-made Caywork agent with 50 free credits and see how clear instructions and the right knowledge keep it on task.

Frequently Asked Questions About AI Agent Instructions and Knowledge Bases

The questions below cover what business users ask most often when an AI agent does not follow its instructions. The answers are short and practical, so you can use them as a quick reference. Each topic is covered in more depth in the sections above.

Why Does My AI Agent Ignore My Instructions?

Most of the time, the instructions are vague or contradictory, the agent has too many jobs, the conversation has run long, or the knowledge base is missing information or contains conflicting rules. Fixing these setup issues solves the majority of cases. On Caywork, starting from a ready-made agent built for one job avoids many of these problems from the start.

What Is an AI Agent Knowledge Base?

It is the set of documents and facts an agent can look up while it works, such as price lists, policies, and FAQs. It gives the agent knowledge about your business that general AI models do not have. A good knowledge base is focused, current, and free of conflicting versions.

What Is the Difference Between Custom Instructions and a Knowledge Base?

Custom instructions tell the agent how to behave: its role, tone, steps, and rules. The knowledge base tells it what it knows: facts and reference material. Keep rules in the instructions and facts in the knowledge base so the two never compete.

What Are Guardrails for AI Agents?

Guardrails are limits that keep an agent within safe boundaries, such as restricting which tools and data it can access or requiring a person to approve important actions. Unlike instructions, which an agent can misread, guardrails cannot be ignored. They also keep the blast radius of any mistake small.

Can I Set Up AI Agents Without Technical Skills on Caywork?

Yes. Caywork offers ready-made agents you can use right away and a visual builder for creating your own without code. New accounts get 50 free credits with no card required, so you can test an agent on real work before committing.

Final Thoughts

An AI agent that ignores its instructions is rarely broken. In most cases it is working from rules that are unclear, knowledge that is incomplete, or more freedom than its job requires. Clear custom instructions, a focused and current knowledge base, and sensible guardrails solve most of these problems before they ever reach a customer. Getting this right matters beyond daily convenience: Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, citing rising costs, unclear business value, and weak risk controls. The teams that succeed treat setup as part of the job, not an afterthought.

If you want to put these ideas into practice, Caywork gives you a simple place to start. You can browse ready-made AI agents built for specific business tasks, or build your own in a visual builder with clear instructions, your own business knowledge, and access limited to what each job needs. With 50 free credits and no card required, you can test an agent on real work, adjust it, and see for yourself how a well-set-up agent stays on task.