Agents Can’t Work From Fragments

My favorite movie is Memento.

The movie revolves around Leonard, a man who suffers from anterograde amnesia and cannot form new memories. Throughout the film, he relies on photos, notes, tattoos, and instructions to understand what happened before, what matters now, and what he should do next.

Every time Leonard acts, he is reconstructing the situation from whatever his past self left behind. The notes he creates act as the memory he cannot carry himself. They are how he connects the moment he is in to what happened before.

That is increasingly how I think about AI agents.

An agent can write, reason, summarize, search, use tools, draft emails, analyze data, and execute steps in a workflow. But every action it takes depends on the context surrounding that action.

What is true right now? What changed? Which source should it trust? What is it allowed to do?

If that context is reliable, the agent can be useful.

If that context is missing, scattered, stale, or trapped in places the agent cannot access, the agent is forced to act from fragments.

And acting from fragments is where things break.

The context is scattered.#

Take a normal work moment: a customer call is coming up, and someone needs to prepare the account context before the meeting.

The agent needs the basics: what the customer cares about, what happened last time, what was promised, what changed internally, and what should happen next.

Most teams already have that information somewhere.

The problem is that “somewhere” is doing a lot of work. It might be in a CRM, a Slack thread, a doc, a meeting transcript, a project board, an email chain, a previous AI chat, or someone’s memory.

A human can often survive that. We know who to ask. We remember the nuance. We can sense when a task title is outdated. We can read between the lines.

An agent does not have that social map. If the context is not carried by the workspace, the agent either guesses, stops, or pushes the work back to a human.

The agent has to verify what is still true#

So whenever the agent has to get work done it first has to answer a more basic question:

Which facts can it still trust?

Was the last customer complaint resolved, or only acknowledged? Did the product team actually ship the fix, or only discuss it? Is the task board current, or did the plan change in a call? Is the latest pricing in the CRM, the email thread, or the deck someone sent yesterday?

A human usually resolves this without noticing. We use memory, instinct, and informal context to decide what to trust.

For an agent, that judgment has to come from the system.

Before it can draft the agenda, suggest talking points, or write the follow-up, it has to know what version of reality it is working from.

If it has to ask you to paste in the latest context, it is not really working from the workspace.

The current workspace still hands the work back to humans.#

This is why adding an agent to an old workspace is not enough.

A workspace built for humans can get away with being incomplete, because humans carry the missing context themselves. A workspace built for agents cannot.

This incompleteness is the moment of failure for the agent, leading to a half-finished task.

If the agent gives you a draft but cannot update the task, CRM, doc, or follow-up, the work still lands back on your desk.

The workspace can no longer be only a place where humans look at work. It has to become a place agents can read from, write to, and be checked inside (e.g., a unified data model, explicit status tracking, and automated source prioritization).

In essence, the new workspace must become the agent’s reliable set of photos, notes, and tattoos, ensuring it never acts from fragments again.

Humans still set direction, judge quality, approve important actions, and carry accountability.

But agents need the workspace to carry enough of the facts for them to act usefully.

So my hot take is that maybe the bottleneck for AI agents is not intelligence.

Maybe it is the workspace they are forced to work from.