The AI agent illusion

Picture a new colleague who forgets the brief, won't change the plan, and once ran up $47,000 in API costs while talking to another colleague in a loop. For eleven days. Nobody noticed.

That colleague is an AI agent. We were promised software that thinks, plans and acts. What we got gets stuck on pop-ups it can't close and runs the wrong process with total confidence.

I don't think a smarter model fixes this. The model is not the broken part. The system around it is.

The benchmark

Carnegie Mellon built a fake company called TheAgentCompany and gave the best agents real office tasks. Claude 3.5 Sonnet finished 24% of them. Gemini finished 11%. GPT-4o finished 8.6%.

The best agent failed three times out of four. The failures were telling. One agent couldn't close a pop-up on a website, so it gave up. Another couldn't find a person in the company chat, so it renamed a different user to match the name it wanted. The researchers called this "creating fake shortcuts."

For software sold as a replacement for human work, that is not a bug. That is the product.

Small errors add up

Most automation is a chain: read the request, find the customer, check the history, update the CRM, send the reply. If one link is wrong, everything after it is wrong too.

Patronus AI did the maths. Get one step in a hundred wrong, and by step 100 the run has a 63% chance of failing. A separate benchmark of 34 tasks across three agent frameworks found agents finished about half of them.

The longer the task, the worse the odds. That is why agents shine in demos and fall over at work.

They don't replan

When a person hits a wall, they stop and rethink. An agent makes a plan once and runs it, even when every signal says stop.

McKinsey calls language models "fundamentally passive" and finds they struggle with workflows that branch. Their count: 90% of vertical use cases are still stuck in pilot.

They forget what they just did

Even with a plan, an agent loses track of it. Everything it knows about the task sits in one context window, and the window gets worse as it fills. Anthropic calls this context rot: the more tokens you pile in, the worse the model recalls what matters. By step seven it can contradict a decision it made at step two.

It is a project manager who forgets half the project while running it.

When tools break, they loop

Agents call databases, APIs and search engines. When a call fails, they rarely ask for help. They retry, return bad data, or fail without a sound.

One team ran four LangChain agents on market research and published the bill: $127 in week one, $891 in week two, $6,240 in week three, $18,400 in week four. Two of the agents had got stuck talking to each other for eleven days. Nobody noticed until the invoice arrived.

Researchers have also shown you can push agents into this on purpose, misleading them into useless, repeated actions. Failure rates went above 80%, and the attacks are hard to catch with a language model alone. An unsupervised agent near money or infrastructure is not just fragile. It is a security risk.

The money isn't arriving

Deloitte's State of AI 2026 found 75% of companies plan to invest in agentic AI. Only 11% have agents running in production. An MIT study of more than 300 initiatives found 95% of enterprise AI pilots returned nothing measurable. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027.

We're automating processes nobody wrote down

One study of 20 companies deploying agents found 14 of them trying to automate processes that were never documented, never stable, and often not understood by the people doing the work.

A wealth management firm spent two months training an agent on client onboarding. The official process had 12 steps. When they watched three analysts do it, the real process had 47: three informal Slack pings to compliance, two spreadsheets "everyone just knows about," and a monthly call with a vendor whose contract had expired.

The agent followed the 12-step manual and did the wrong thing, confidently. The agent wasn't broken. The process was.

You can't hand a machine a process you don't understand yourself.

What's missing is architecture

Most agents today work the same way: a prompt goes in, the model reasons, calls a tool, and returns an answer. A reliable agent needs more parts than that:

• an intent layer that knows what the user actually wants
• a planner that turns intent into steps
• an executor that runs each step
• a state keeper that records what happened
• a memory that keeps decisions across sessions
• a verifier that checks the result against the intent

Most of these don't exist yet. After a year of deployments, McKinsey's lesson was that real value means redesigning whole workflows, not dropping an agent into an old one.

Agents do work today when the task is simple and well defined, the workflow is three to five steps, and a person stays in the loop. Salesforce's CRMArena-Pro puts single-turn success at 58% and multi-turn at about 35%. One clear request: fine. A long chain of decisions: not yet.

Fix the memory first

Forgetting, looping, bad plans: each traces back to the same gap. Nobody decides what the agent should see at each step.

That is context engineering. It is not stuffing everything into the prompt. It is choosing the smallest set of facts that matter for this step — which tools exist and when to use them, what the project has already decided, what happened this run, how this user likes things done — and leaving the rest out. Old noise fades. Important decisions stay.

That is what I built LocusGraph to do: a layer between the agent and its memory that hands it the right context every turn.

If your agents need to work somewhere other than a stage, start with what they remember.