AI Agent Workflows: Productivity Gains and

There's a workflow making the rounds in AI power-user circles right now, and it's worth your attention — not just because it's productive, but because it asks something of you that most tutorials quietly skip over. That thing it asks for is access. Broad, local, "go look at my whole file system" access.

Nate Jones, who runs the AI News & Strategy Daily channel, published a short video recently laying out how his daily AI practice has shifted in the past few weeks. The core move: instead of manually feeding documents into a chat interface, he instructs Codex to browse his local file system, locate relevant files by description rather than filename, copy them into a clean working folder, and then execute long-running tasks against that assembled context. He reports handling 30,000-to-50,000-word document jobs this way — complex drafts, spreadsheets, code — with a consistency he hasn't found in other tools.

That's a genuinely useful technique. It's also a technique with a threat model, and Jones doesn't spend much time on it. So let me.

What "look at my file system" actually means

When Jones tells Codex to scan his local file system and find files by natural-language description — "it's about X, I made it around Y time, find it" — he's granting an AI agent navigational access to his entire machine. Not just a designated folder. The whole thing. And then he's trusting it to copy the right files into a working directory and ignore everything else.

In practice, this means the agent is traversing directories it was never explicitly pointed at. It's reading filenames, likely file contents, and metadata across your system to pattern-match against your description. That process happens on your local machine with Codex, which is meaningful — the data isn't being routed through a third-party server in the same way a cloud-based chat interface would. But "local" doesn't mean "contained." It means the agent has broad read access to whatever your user account can reach, which on most people's machines is: everything.

Before you replicate this workflow, ask yourself a few questions. Does your work machine share a filesystem with personal documents, financial records, or health data? Do you have credentials, API keys, or auth tokens stored in plaintext files anywhere — a .env file, an old notes document, a config you set up years ago and forgot about? Are there files belonging to clients or employers that carry confidentiality obligations? Because when you tell an AI agent to "look around and find what's relevant," you are relying on its judgment — and the accuracy of its search — to determine what it touches and what it copies.

Jones says it "inevitably finds" what he's looking for. That's the pitch. What he doesn't address is what else it finds, reads, and processes in the course of that search.

The tradeoff is real, and it might be worth it

None of this means the workflow is bad. It means the workflow has a cost, and you should go in with your eyes open.

For someone working with proprietary but non-sensitive creative or professional documents, on a dedicated work machine with sensible file hygiene, this is probably a manageable risk. The productivity gains Jones describes are significant enough that a lot of people will reasonably decide they're worth it. The ability to run a 50,000-word editing job, simultaneous multi-threaded drafts, and complex cross-document synthesis without manually curating every file is a real capability unlock.

For someone whose machine also holds client contracts, medical information, personal financial data, or anything that lives under a legal or ethical confidentiality obligation — the calculus is different. Not necessarily a hard no, but a reason to set up a dedicated machine or partition for this kind of work before you start.

The minimum viable security practice here: create a designated input folder, put only the files you intend to share into it, and point the agent at that folder — not at your system. You lose some of the "I don't have to keep track of my files anymore" magic, but you keep control over what the agent actually sees. If that tradeoff sounds annoying, that's because it is. Security is often inconvenient. That's not a reason to skip it.

The prompting shift is the more interesting story

Separate from the file system question, Jones describes something that feels like a genuine inflection in how people are working with AI agents. His framing:

"Help me to define the shape of this task first and then once we define it then we can go execute it agentically."

That's a different relationship with these tools than what most tutorials teach. The conventional advice — still circulating, still useful for simple one-off tasks — is to engineer a tight prompt, hand it to the model, and evaluate the output. Jones is describing something more like collaborative scoping: use the model's reasoning capacity to figure out what the task actually is before you unleash it to go do the task.

He traces this shift to recent model updates, specifically referencing what he calls "Claude 4.7" and "Claude 5.5" — naming conventions I can't independently verify against Anthropic's current release catalog, so take those version numbers as Jones's own terminology rather than confirmed product names. He also mentions trying this same workflow with "Claude Code" and something he refers to as "Claude Co-work" — the latter isn't a product name I recognize from Anthropic's lineup, and it may be a misstatement or an informal label he's using for something else. He finds Codex outperforms both for this specific file-based approach, which he attributes either to Codex's origins as a code-generation tool trained extensively on structured file repositories, or to what he characterizes as compute constraints on Claude's end — though that's his read of the situation, not an established fact.

The architectural argument is plausible without being the whole story. Codex has deep familiarity with navigating structured file sets from its training on public code, but "built for repository navigation" oversells the intentionality — it's more that those skills transfer usefully to text files in folders. Whether the performance gap is architectural, resource-related, or just where the tools happen to be right now is genuinely unclear, and Jones is honest about that uncertainty: "I'm not sure," he says, which is a refreshingly non-confident thing to say in a space full of confident declarers.

What the collaboration shift costs you

The prompting evolution Jones describes — from instruction-giver to collaborative task-shaper — is worth taking seriously. But it also reframes a question that deserves more attention than it usually gets.

When you use an AI agent to help you define the shape of a task before executing it, you're not just delegating execution. You're delegating part of the framing. The model's suggestions about how to scope your task, what questions matter, what "good" looks like — those suggestions are products of its training, its design, and the values baked into it by a company with its own interests. Most of the time that's fine. Sometimes the framing it offers is better than what you'd have come up with alone.

But the more seamlessly these tools shape your work before you even notice they're doing it, the harder it gets to locate where your thinking ends and the model's begins. Jones closes with a line I find genuinely clarifying: "I don't need any particular team to win here. I just want to get more efficient at working." That's an honest and reasonable position. It also implicitly accepts that the teams building these tools are making choices — about data handling, about agent permissions, about what the autoreview system flags or doesn't flag — that shape what "efficient" looks like in practice.

You're not just adopting a workflow. You're adopting someone else's assumptions about how work should go, embedded in a system you didn't build and can't fully inspect.

That's worth knowing before you hand over the keys to your file system.

Rachel "Rach" Kovacs covers cybersecurity and privacy for Buzzrag.