
The context layer has a clutter problem
In 1953, a British engineer named Colin Cherry set out to explain a phenomenon he called the cocktail party effect: the way humans can tune out everything but the voice they’re listening to in a crowded room. He gave a group of participants headphones that played two different speeches at the same time, one in each ear, and asked them to repeat one of the speeches out loud as it played.
The participants recited the speech easily, but Cherry was interested in what happened to the other one. Was it being processed in the background? Could participants recite parts of it too? When he asked, he learned they hadn’t even noticed that it had switched from English to German partway through.
Cherry’s finding laid the foundation for the first theory of human attention, which has held onto its core premise through decades of scientific iteration: human minds don’t process as much information as possible and then decide what matters. They set a goal that determines what matters—like reciting a specific speech—and filter out everything not relevant to it. This filtering is what lets new parents sleep through a thunderstorm and wake at the first sound of a cough, or employees leave a meeting with only the information relevant to their team. It’s the reason people fail Daniel Simons and Christopher Chabris’s famous selective attention test (go ahead: try it for yourself). We don’t notice it happening, but nearly every judgment call we make starts with this process of selectively ignoring any information that’s not crucial to our goal.
And LLMs are bad at it.
—
This year, a growing number of companies have entered the business of building an AI “context layer”—a broad term for any tool that informs or decides what AI agents read in order to succeed at tasks. In the first week of November alone, Notion, Box, and Snowflake all announced plans or products that give agents business context in different ways. These launches are, at least in part, responses to a disappointing first year for agentic deployment, in which agents made some employees faster but failed to make companies any better. The market’s clear POV is that giving those agents more business context is part of the solution to those disappointments.
The products describing themselves as part of the context layer range from semantic layers and permission-aware retrieval to intranets, connectors, and ontology programs. These technologies serve different purposes, but they almost all share the same foundational assumption that if you give an agent enough information about a company, the model will figure out how the work gets done.
That assumption is wrong. Documents, meeting notes, databases, Slack messages, and every other source the context companies mine capture the coordination of work that needs to be done or they capture the outputs of the work after it happened. These traces give models valuable clues about how work gets done, but that is all they give them: clues. And before companies invest millions of dollars into technologies that help generalized models approximate how work happens, it is worth asking if high-volume clue retrieval that never looks at the work itself is the best we can do.
It’s easy to understand why context layer businesses are focused on building products that mine information companies already have. The SaaS era produced systems of record for almost every type of business data imaginable: HRIS tools to keep track of people, CRMs to monitor leads and relationships, project management tools to track employee to-dos. There’s a tool for cataloging almost everything except for the way actual work gets done. For decades, that gap has persisted because the information was virtually impossible to capture at scale.
Even in a midsize company, work happens across dozens of teams and tools at every second of the workday, and only a fraction of it gets documented as how-to guides and process maps, even when you make documentation a priority or hire consultants to do it. To accurately capture and maintain a documentation database of how a whole organization works would take thousands and thousands of hours.
But without those records, models are left to reverse-engineer how work gets done from only what it leaves in its wake.
—
Agents are notably bad at that reverse-engineering, but not for reasons that more access to existing business context will fix. Unlike Cherry’s study participants, transformers don’t have the choice to completely tune out information they don’t need: every piece of information in front of them gets read and weighed against everything else in the window. Direct them with a goal, and they’ll weight what’s more relevant more heavily, but even newer models are easily misled by things that don't matter. To a model, a goal is just more text in the same window as everything else, not a filter applied before the reading starts.
The companies building these context tools know this, so they’ll build the filters themselves. This can be done in several ways, but the most common approach is some variation of search: an agent queries company files and loads in only the closest matches.
To understand how this mechanism works in practice, imagine the way a rep prepares for a renewal call. If they’re experienced, it’s likely they’ll ignore existing company documentation about how to handle renewal calls and go re-read something that, on paper, has nothing to do with the renewal, like a Slack message from the account champion, the customer’s main advocate for the product, mentioning they’re interviewing for new roles. The rep does this because they know that the guidance won’t tell them anything they don’t already know, and what matters in practice is that champion departures pose an existential threat to contract renewals, even for loyal, satisfied customers.
An agent preparing for the same call, even with a sophisticated search filter in place, will undertake a very different process. First, it will retrieve whatever messages, documents, or other artifacts most resemble the task query. In doing so, it will return the sales team’s guidance on renewal calls, last year’s pricing deck, and basically everything a good rep would ignore. What it won’t retrieve is the Slack message about the account champion’s job search because nothing in that message resembles the query about renewals, and even in a smart system that did catch it, that result would arrive along with a dozen others that look far more on-topic.
By filtering this way, context tools will surface exactly the kind of information research shows trips models up most often: material that looks relevant but isn’t.
Last year, Chroma tested eighteen models on their ability to answer questions surrounded by two types of filler content: content that was totally irrelevant to the question and “distractor” content, content that was closely related to the question but didn’t actually answer it. They found that even one distractor measurably lowered performance, a pattern Google researchers also documented in 2023. Since most companies possess hoards of documents and messages related to any given task, the best matches for almost any query are likely to be distractors.
Naturally, retrieval will get better with smarter queries, rerankings from second models, better ontologies, and more. There’s no doubt in my mind that these improvements will arrive, but a better search is still just a better guess about which traces matter, and it will not close the gap between agents and humans on its own. In the case of the renewal call, the information the agent needed to prep for the call did already exist in Slack, where a human counterpart would look for it. But what that agent lacked, and what it will always lack when left on its own to deduce how work happens from its residue, is any reason to believe it needed that Slack message in the first place. Better search will fail for the same reason more context does: nothing in a company’s traces says which of them matter. The only type of data that can begin to reveal this, even imperfectly, is high-fidelity mapping of workflows and how work really gets done.
—
We ran a small test of this idea in the fall to understand if workflow data can give models an understanding of how to do work that other contexts cannot. We gave a model thirty business tasks across sales, marketing, and product, each with a judgment element that couldn’t be answered from a documented rule alone, and ran every task under four conditions: the model alone, the model with our internal wiki, the model with step-by-step Scribe recordings of people completing workflows, and the model with both. An LLM judge scored each output on seven dimensions, including specificity, actionability, company relevance, and how well it applied what it knew to the new scenario.
With nothing but the model, the output was mostly long documents with placeholders and generic advice that most employees would hate to read. Adding either the wiki or the recordings improved its performance by more than half, but the workflow recordings beat the wiki on every measure we scored.
The surprise was that the model didn’t benefit from having both. Counterintuitively, adding the wiki to the model with access to workflow recordings actually degraded its performance. More is not always better.

Of course, this is one experiment, run on one model and graded by that same model. But it reflects what, at Scribe, we have long believed will be the secret to unlocking agent performance: because a workflow recording contains only information about how a task is actually done, it is the one thing we can give agents where the filtering work the human did has already been applied. Though it will likely never be possible to capture every contour and nuance of human judgment in how we train agents, by letting agents learn from the post-filter choices of actual humans, we can get closer than ever before.
—
The context layer is going to keep filling up. More connectors, definitions, and indexes are on their way, and each will give agents another set of clues about the one thing companies have never had data for. To be clear, this is better than nothing. What those clues won’t contain is the part of a job that experienced people find hardest to explain and that models cannot infer from documents—decisions about what to ignore. That expertise exists only in the work itself, in the thousands of choices people make every day about what deserves their attention, which companies never catalog or keep. Until companies start capturing that work and recognizing it as information their agents need, each new connector will just be one more voice shouting in a crowded room.





