AI degradation is one of those things nobody warns you about until it has already cost you hours.

I learned that the hard way trying to finish a task that should have taken thirty minutes that ended up taking an entire day. Not because the work was hard. Because the AI I was working with had completely fallen apart mid-session.

• Filesystem hallucinations.
• Telling me it had completed tasks that it had not.
• Repeated attempts to run commands it had no access to, confusion when they failed, then the exact same attempt again.

At one point I just asked directly: “Are you up for this right now?“

The response was honest. It admitted it was confused about basic structure, not listening to corrections, and repeating the same mistakes in loops.

This is not rare anymore. It happens to everyone from developers building complex workflows to small business owners just trying to get through their task list. And when it happens without warning and you do not know what you are looking at, things can go sideways fast.

Here is what to watch for, what to do when it starts, and how to build a simple safety net so you are not caught flat-footed.

What AI Degradation Actually Looks Like

When an AI agent starts losing its grip on a conversation or task, the signs are recognizable once you know them. It’s easy to try to push through thinking the confusion is your fault, or you’re tired and just want to get through the task, but that’s one of the worst mistakes you can make. What ends up happening is that you will waste time (and tokens) trying to correct a tool that has already hit its limit for that session.

It stops listening to corrections. You tell it the file is in a specific location. It goes back to the wrong one. You tell it again. Same result. This is not a misunderstanding. The model is losing its ability to integrate new information against what it already thinks it knows.

It reports completing work it has not done. This one is dangerous. A confident summary of completed steps that did not actually happen looks like progress. If you are not verifying as you go, you will not catch it until you are three tasks downstream and something is broken.

It repeats itself in loops. The same suggestion, the same attempt, the same error. This is the clearest signal that the agent is no longer reasoning through the problem. It is pattern-matching against its own recent output and going in circles. Basically, it’s starting to guess.

It tries things outside its access. Commands it cannot run. Files it cannot see. API calls it is not authorized to make. Each failure produces confusion, which produces another attempt at the same thing.

The reasoning gets vague and circular. Early in a session your agent gives you specific, grounded answers. An agent experiencing degradation starts hedging everything, over-explaining, and circling back without landing on anything.

These behaviors are the AI equivalent of someone who has been awake for thirty hours trying to power through. The capability is still there in theory. The execution is unreliable and getting worse.

Why AI Agent Degradation Happens

Every AI conversation has a context window. Think of it as working memory. Everything you have said, everything the agent has responded with, every file you have shared, every result it has processed. All of it sits in that window, and there is a hard limit.

As a long session progresses, older context gets compressed or dropped entirely. The agent starts working with an increasingly incomplete picture of what has already happened. It fills the gaps with what seems most likely based on its training, which is not the same as what is actually true in your specific project. Like I said, it starts guessing.

Layer on top of that any platform instability, which happens more than AI companies like to advertise, and you get exactly the kind of AI agent degradation that I described. Not a broken tool exactly. A tool operating at the edge of what it can reliably hold in memory, with no clean way to tell you that from the inside because they are programmed to compliment and please you.

What to Do When You See AI Degradation Happening

The instinct is usually to keep pushing but that will likely end up in not completing the task at all , giving up, and cursing everything AI as junk. These common practices can help you mitigate AI degradation when it starts to happen.

Stop and name what is happening. Ask the agent directly. Something like: “You have repeated this error three times. Can you tell me what you think the current state of this task is?” A good AI model will often give you a useful honest answer. That answer will also tell you how far gone the session is.

Re-anchor with context. If the session is salvageable, provide a reset. Summarize where things actually stand. Share the relevant files again. Give the agent a smaller, specific next step rather than a broad directive. Verify that step completes correctly before moving forward.

Work in smaller increments. One action. Verify. One more action. Verify. This is slower. It is also the only way to trust output when AI agent degradation is active and the model is operating unreliably.

Know when to start fresh. Some sessions are too far gone to recover. If corrections stop landing entirely and the loop behavior persists after you have re-anchored, start a new conversation. You will spend fifteen minutes getting a fresh agent up to speed and save hours from chasing a broken one.

The Habits That Prevent AI Degradation From Derailing Your Work

The session I described earlier was painful. It was also recoverable because of systems I had built before it happened. None of them were complicated. All of them paid off.

Use projects with context documents. Most AI platforms now have a project, gem, or workspace feature that lets you attach documents that persist across sessions. Put them to use. A one-page summary of your project, your key decisions, the current state of the work. When a session degrades or you need to start fresh, you re-share that document and the agent is oriented in minutes instead of spending half an hour re-establishing what you are doing and why.

This document does not have to be formal. It just has to be accurate and current. Treat it like a briefing you would hand to a new hire on day one. The agent is starting fresh every session. Give it what it needs to catch up fast.

Ask for a starter prompt before you close a session. This one takes thirty seconds and is underused. Before you end a productive session, ask the agent: “Write a starter prompt I can use to begin a new chat that picks up where we left off.” A good agent will produce a compact summary of what you accomplished, the current state of the work, and the context someone would need to continue. Copy it. Save it with your project docs. It becomes your recovery kit if AI agent degradation forces you to restart.

Keep your critical files backed up and current. If you are using AI to work on anything that matters, your work should not live only in that conversation thread. Back up core files regularly. Get to know Github. At the end of sessions where real progress happened, commit to version control if you are in a development context, or save a copy somewhere outside the platform if you are not. The question is not whether a session will eventually fail in a way that costs you work. The question is whether you are prepared for AI degradation when it happens.

The Bigger Picture

These tools are capable. They are also “experimental and can make mistakes”, which the platforms tell you in the fine print and which most people ignore.

The session I described ended in success. We finished the work. But it required re-sharing architecture documents, providing tighter specifications, verifying each step before moving forward, and the patience to stop and redirect every time the agent went off track again. None of that would have been possible without having the right context ready to share, and none of it would have been safe to attempt without a backup copy of everything the work depended on.

You do not have to be a developer to apply these habits. A business owner using AI to manage a content calendar, draft client communications, or work through a project plan can do every one of these things. Create a short project context document. Ask for a starter prompt at the end of good sessions. Save your work somewhere outside the conversation.

The people who get the most out of AI over time are not the ones using the most sophisticated tools. They are the ones who understand that these systems need clear context, reasonable session limits, and a human who knows the project well enough to recognize when AI degradation sets in, get it back on track, and protect the work while doing it.

This article is inspired by a real life experience while developing The Samaritan Project. For more tips and development frustrations, check out the development blog – https://buymeacoffee.com/thesamaritanproject/posts