CERTAINCE Logo

Where AI Agents Fail in 2026 and What Helps

AI Content · 14 min

Illustration: an AI agent draws a free path between fixed golden gates that stand at the decisive points

Drew Breunig has been writing about context engineering since 2025: what information a language model has in front of it when it works on a task. In October 2026 he gave a short talk on the state of the field at AGNTCon, and on 24 October he published the slides and a transcript. He walks through the failures that arrived with the large context windows of 2024, the fixes that have held up, the new failures that 2026's long-running agents bring, and what he expects for 2027. His conclusion applies at every stage: whoever works with AI decides what a fixed program handles and what the model may decide for itself.

We work with AI agents every day in our own company. Our own agent control system proposes tasks from emails and selected WhatsApp conversations, agents work through them, and a person approves anything that leaves the company. The agents draft replies, write and review code, and maintain our knowledge store. Our founder Sebastian Behler owns the store. He sets new rules and decisions. The agents write routine updates without a sign-off for each one. This article follows Breunig's outline and sets next to each stage what we saw in that work and what we changed as a result. Where we have nothing of our own to add, we say so. What context means in a business and who maintains it is covered in context management for business.

First, a large context window was supposed to be enough

In February 2024 Google introduced Gemini 1.5 Pro, a model that can process one million tokens at once. A token is a piece of a word, the unit in which language models count text. According to Google, one million tokens is more than 700,000 words; the first Gemini generation had handled 32,000. The context window, the amount of text a model takes into account for one answer, had grown many times over. That raised a convenient hope: if everything fits, nobody has to choose anymore.

In June 2025 Breunig described why that does not work. A long context has room for more information and just as much room for more errors. He distinguishes four kinds:

  • Context poisoning: an error gets into the context and is referenced again and again. Google's Gemini 2.5 Pro spent many hours of a Pokémon game looking for tea that does not exist in that version of the game. That case and one of our own are described in AI hallucination or poisoned context?.
  • Context distraction: the history grows so long that the model orients itself more by the history than by what it learned in training.
  • Context confusion: information that has nothing to do with the task influences the result.
  • Context clash: new information, tools or instructions contradict what is already in the context.

Carolina Camassa and Derek Shiller measured how strong distraction is in "Do as I Say, Not as I Do" in 2026. They gave 13 models a clear instruction, such as translating every question into Russian. They then placed examples in the history in which the assistant did something else and answered in French. With each additional example, the models followed the pattern more often and the instruction less often. All 13 were affected, to very different degrees. Neither a model's size nor its results on the usual benchmarks predicted how long it held out.

In an office, the situation is familiar from long chats. If you have a draft revised five times in the same conversation, you have placed five versions you did not want in the tool's history. Whether that measurably worsens the next answer is not something the study examined; it worked with prepared conversations. Google observed something related in a real agent. The Gemini 2.5 technical report says that the Pokémon-playing agent tended to repeat actions from its history instead of developing new plans once that history grew significantly beyond 100,000 tokens. The report itself calls this an anecdotal observation. The model could process more than a million tokens.

Against the other two kinds of failure we have rules. Breunig switches off the memory feature in ChatGPT and Claude because it mixes notes from his work with notes from personal projects. Our agents' automatic memory stays off as well. Business facts are kept instead in files, with a date and a source, where someone can check them. Against clashes, each rule has one file responsible for it, and other files point there instead of repeating it. Where two files could still contradict each other, it is settled which one takes precedence. For the copy on this website, for example, one file decides on wording and another on prices. An agent that finds a contradiction reports it instead of quietly picking a side.

The fixes have stayed the same since 2025

Also in June 2025, Breunig described six fixes. In his talk he notes that they have stayed the same even as new failures appeared. We use all six:

  • Augmenting: the model gets the information it lacks for exactly this task, for example from a document search or a work instruction. Before one of our agents drafts a reply to a client, it reads the email history so far and the project notes.
  • Tool loadout: the model sees only the tools and instructions it needs for the task. Breunig tells of a company whose agents suddenly performed badly because 600 work instructions were loaded at every start. With us, an agent first sees one line of description for each work instruction and loads the full text only when the task needs it.
  • Quarantine: a subtask goes to a separate agent that starts without the history so far. When one of our results needs an independent review, an agent from the other provider usually does it. It sees the brief and the result, but not the conversation in which the result was produced (how we check AI-written code).
  • Pruning: whatever is no longer needed leaves the history. Claude Code and Codex partly do this on their own. When our agents check a web page, they use a command-line tool that reports only the result, so that not every snapshot of the page ends up in the history.
  • Offloading: information is stored in files or a database and read again when needed. With us, every durable fact sits with a date and a source in exactly one file of our knowledge store.
  • Summarizing: a long history is replaced by a summary. When one of our agents passes a task on, it hands over a few sentences: the current state, the next step and any open decisions. The section on 2026 shows why there are so few.

If you build agents yourself, Breunig's transcript includes short example code for four of the six fixes.

The names, by contrast, have changed quickly. In 2023 this work was called prompt engineering, in 2025 context engineering, and in 2026 people speak of harness engineering, after the framework program an agent works in. The job stayed the same: giving the model the right information. What changed is who assembles it. With each step, fixed program code takes over more of it and the model itself less.

2026: agents that work for hours

In 2026 the model providers promise something other than a larger window. The models are meant to operate their harness themselves, manage their own context and stay on one task for many hours. In an experiment Anthropic described in October 2026, agents had six hours. With us, a coordinating agent, one that distributes tasks to other agents and brings their results together, may work for up to four hours at a stretch. Why not longer is explained further down. Breunig sees two new kinds of failure in this development.

Fighting the weights

Breunig calls the first one "fighting the weights": working against the weights the model learned in training. A model that learned to operate a harness in particular environments brings their habits along and therefore sometimes does not do what you ask. As evidence, Breunig shows instructions from Claude Code, Anthropic's coding tool. The request to run independent tool calls at the same time rather than one after another appears there in many places, sometimes repeated word for word. His rule of thumb: if you find repetition and capital letters in an instruction, you are probably fighting the training.

Anyone who writes with AI knows one example: the dash. On 14 November 2025, Sam Altman, OpenAI's CEO, reported as a small but happy win that ChatGPT no longer uses em dashes if you ask for that in your custom instructions. Before, it had not followed that reliably. Our writing rules have also banned dash constructions since 31 August 2026. The rule appears in several instruction files. In addition, a check script counts the dashes in every article before it is published, and none is allowed.

We apply the same principle to rules with more at stake. That agents do not send emails is in their instructions and is also technically ruled out: their access to our mailbox has no permission to send. Our AI provider's ready-made Microsoft 365 connection could send. So its tools for sending emails and Teams messages are switched off in our settings, and the agent uses our own mailbox tool with fewer rights. When an agent reads one of our live PostgreSQL databases, the tool opens a transaction that only allows reading. A rule you have to repeat in an instruction belongs in a permission or in a check run by a program.

Getting stuck

Breunig calls the second kind of failure "context inertia". An agent fixates on a hypothesis, a subtask or a goal and keeps pursuing it long after it should have turned back. In October 2026 several papers on it appeared in quick succession:

  • In its report on experiments testing whether AI models can replicate themselves without access to a datacenter, Anthropic describes a harness called Memento. It is meant to catch a common failure in long runs: the agent holds on to a wrong idea, for instance about the cause of a bug, instead of backtracking. Memento therefore periodically starts a fresh agent. It takes over its predecessor's workspace and short notes, but not its history. In one of the experiments this happened every hour and a half.
  • A team led by Ziyang Cai at Microsoft Research shows in "Fork-and-Flush" that runs of the same agent on the same task often stall at very different levels, even with a lot of additional compute. As a fix, new branches take over the existing workspace and start with an empty history.
  • Shangyin Tan and colleagues from Berkeley and Stanford observe in "Scaling Discovery Requires Forgetting" that agents anchor themselves to their own earlier attempts and make fewer and smaller changes. They therefore periodically discard everything the agents have accumulated and restart from the best result so far. In their experiments this produced up to 15% better results than agents working in parallel and up to 39% better than a single agent.

Breunig notes that none of these papers invents a new fix. All of them rely on quarantine, offloading and discarding the accumulated history. His first question to anyone who complains that their agent has got worse is therefore: how long is your context? From about 200,000 tokens he pays attention himself and starts fresh before a bigger task.

With us, this failure looked like this. From 11 to 13 October 2026, our coordinating agents spent days on tools meant to safeguard our checks while finished changes waited to be merged. On 13 October we also found a handoff plan that had grown to 1,035 lines and had been read 477 times. It was meant as memory between sessions. We read it as the same inertia Anthropic describes, spread over many sessions. The same day, we changed the rules:

  • A handoff is a few sentences: the current state, the next step and any open decisions. It is replaced at every handoff, not extended.
  • It sits where the next session starts, in the task description.
  • Each task gets its own session.
  • A coordinating agent ends its run with such a handoff as soon as its task is merged, and after four hours at the latest. In long runs, hours were lost to the tool summarizing the history.

The work itself is not lost. Code sits in version control, facts sit in the knowledge store. Only the history is discarded, as with Memento. We have not measured how much time this saves us.

2027: memories that contradict each other

For 2027, Breunig predicts the next problem. Personal agents such as Hermes or OpenClaw keep memories for months and write their own work instructions. Breunig has been running a Hermes agent himself for six months and observes its instructions filling up with odd one-off fixes. The agent solves one task one way and a similar one later another way, and instructions and memories start to contradict each other. What is needed, he says, is strategic forgetting, and nobody has yet figured out how to do it. What these agents can do today and what a business should settle before connecting one is covered in our comparison of OpenClaw, Hermes, Grok Bot and OpenAI dots.

We see the beginning of this in our own knowledge store, which agents update every week. In October 2026 we found a file there that cited our own articles as proof of its statements and had turned a balanced argument into a fixed rule. That case is described in AI hallucination or poisoned context?.

So forgetting is planned work for us. Since 10 August 2026 an agent has compared the store with emails, chats and open tasks every week. That includes fixed steps:

  • A check script flags entries on ongoing work whose newest date is older than 14 days, and all others after 42 days. A flagged entry is checked against its source before it gets a new date.
  • The file every agent reads first stays at around 100 lines. Long entries hand their older history off to log files that only those who need them read.
  • Every three weeks a clean-up pass follows. It merges duplicate entries into the file that fits them most precisely and removes only superseded status and questions that were already answered before the previous pass.

That does not solve the problem. We have not measured whether three weeks is the right interval, and what counts as superseded remains a judgment call in every pass. For a business, one clear rule still follows: whoever gives an agent a memory also needs a fixed date on which a named person tidies it up, and a rule for what may be removed.

What a program settles and what the model decides

Breunig closes with a slide he says he made a year and a half ago for corporate training. The output of a language model depends on two things: the data it was trained on and the context it receives for the task. If you want better results, you can use a larger model, which is expensive, let it think longer, which is slow, or improve the context. Only the last is cheap and fast. For Breunig, the job is deciding where a fixed program works and where the model may drive. Every advance in the models moves that line.

In our work we draw the line with one question: can the rule be checked the same way every time? Then a program checks it, or a missing permission makes the tool refuse the step, whatever the instruction says. If it needs judgment, the model drafts, and a second model or a person reviews. With us, the split currently looks like this:

  • A program or a missing permission decides who may send emails, whether a text contains dashes and whether a code change passes its tests together with the current state.
  • The model reads email histories and drafts replies, writes code and, as a second agent, checks whether a draft claims more than its sources support.
  • A person decides what leaves the company and which version of a piece of software goes live, and sets new rules and decisions for the knowledge store.

Review this split when you change models. A new model follows some instructions the old one ignored, as ChatGPT has followed the instruction on dashes since November 2025. It also brings new failures, as the long runs of 2026 have shown.

Where this line runs in your business is what we settle in the setup: which sources an agent reads, which permissions it gets in your existing programs and who approves what. Rules that a written instruction cannot carry, we build as software. Book a free AI assessment

Inquiries