Research
How much of an agent’s work is spent finding files
What public agent traces reveal about file access, why counting operations and measuring time tell different stories, and what enterprise teams still cannot measure.
Of all the work an AI agent does on a task, how much of it goes toward locating and reading files, and how much goes toward the task itself?
There is no reliable percentage for enterprise environments yet.
Public research gives us several pieces of the picture. Machines are already significant readers of documentation. We can observe how coding agents interact with tools. And enterprise retrieval benchmarks show that finding complete information across large document collections remains difficult.
What these studies cannot establish is how that activity translates into requests against production enterprise file systems. That distinction matters: an agent’s workload cannot be understood from task completion, tool latency, or storage throughput alone.
Machines are already reading the documentation
GitBook measured a week of traffic across its hosted documentation sites from 27 April to 3 May 2026. Across 61,194,835 page views, AI agents accounted for 51.8% of intentional documentation reads, compared with 48.2% by humans. Crawler traffic was excluded from that comparison. About 22% of agent traffic had no identifiable bot name.
A documentation site is not an enterprise file estate. Its traffic cannot be assumed to represent private file shares, corporate repositories or internal document systems. What it does show is that the change in readership is measurable: in this environment, machines already access documentation about as often as people.
Inside an enterprise, identifying those readers is more complicated. Some agents operate under dedicated service identities; others authenticate using a person’s credentials. Whether their operations can be attributed to an agent depends on the identity model and the available telemetry. Even a basic question can therefore be hard to answer: how much of an organisation’s file activity is generated by machines rather than people?
One task can generate many operations
An agent working on a document task may search for candidates, inspect metadata, open several files, follow references and discard irrelevant results before continuing. Each step can generate another request.
The University of Washington’s TraceLab project provides a detailed view of tool activity in coding-agent sessions. Researchers collected approximately 4,300 sessions involving 433,000 tool calls from Claude Code and Codex usage. Each user request generated an average of 10.8 tool calls across 8.8 autonomous model-tool cycles. Across all calls, 76% were shell or command executions, 11% file edits, and 9% file reads or searches.
Coding agents are not representative of every agent application, and shell commands can hide the filesystem operations they trigger. These figures cannot tell us how many file requests a typical enterprise agent generates. They do establish a useful distinction: one user request is not necessarily one infrastructure operation.
Counting operations and measuring time give different answers
TraceLab also measured how long individual tool calls took. In the research team’s aggregate summary, calls shorter than one second represented approximately 61% of tool calls but only about 1% of total tool execution time. Roughly 4% of calls lasted longer than a minute and accounted for about 85% of tool execution time.
By count, short operations dominate. By elapsed time, long operations account for most of the total. Both measurements are correct; they describe different aspects of the workload.
A dashboard focused on slow requests can report healthy file-access latency while large numbers of individually fast operations continue to arrive. Those operations still require processing, even if none is slow enough to trigger an alert. That does not establish that they are expensive. It establishes that individual latency alone is insufficient to characterise aggregate impact.
A workload made of fast operations is easy to miss when monitoring is designed to catch slow ones.
Finding the right file is not the same as reading a file
A successful file read tells us that the system returned the data requested. It does not tell us whether the agent found the information needed to complete the task.
In HERB, published at EMNLP 2025, researchers evaluated deep search across 39,190 synthetic enterprise artifacts. The strongest agentic RAG methods achieved an average benchmark score of 32.96, with evidence retrieval identified as the main bottleneck. Of the benchmark’s 1,514 questions, 699 were intentionally unanswerable, so the score reflects more than locating evidence: it includes appropriate handling of insufficient information.
In AGORA, published at EMNLP 2026, researchers examined archive-grounded reasoning across 9,664 authentic documents, 372 million tokens and 362 questions spanning eight domain collections. The strongest of eight evaluated models achieved 59.4% accuracy.
The benchmarks have different tasks and scoring methods, so their numerical results must not be compared directly, and neither measures file-system requests, storage latency or infrastructure cost. What they show is that retrieving and using complete evidence from large document collections remains difficult. An agent may complete every file operation successfully and still fail to find the evidence it needs, or produce a correct answer after many intermediate operations.
Retrieval success and infrastructure activity are different measurements. Knowing one does not tell us the other.
The number nobody has
We have public measurements of machine readership, agent tool activity, tool-call latency and retrieval quality. What we do not have is a representative measurement connecting those observations inside a production enterprise file estate.
How many file-system requests does an agent task generate? How much of that activity is metadata rather than data? How does the request pattern change when tasks run concurrently? What happens to the people and applications already using the same infrastructure?
These are infrastructure questions, not model-quality questions. An agent trace records operations visible to the runtime. A retrieval benchmark evaluates whether sufficient evidence was found. A storage dashboard measures the workload observed by the infrastructure. None of them alone connects the task to its underlying file activity and resource consumption.
What enterprise teams can measure today
Enterprise teams do not need to wait for a new benchmark to establish useful baselines:
- Metadata request rates. Track directory listings and attribute lookups alongside data reads and throughput.
- Request distribution. Measure how access is distributed across files, directories, applications and time windows.
- Concurrency. Identify overlapping workloads and, where identities permit, distinguish machine-generated requests from human activity.
- Latency distribution. Track request counts and latency percentiles by operation type, rather than relying on averages alone.
Existing infrastructure offers some of these measurements, although attributing requests to individual agent tasks may require additional instrumentation. Storage telemetry does not establish whether an agent found the correct information; that requires application-level measurement. What these baselines can show is how the infrastructure behaves as a new class of workload arrives.
Where this leaves the question
Public evidence supports three observations: machines are significant documentation readers; agent tasks can involve multiple tool interactions; and retrieving complete evidence across large document collections remains difficult.
What remains insufficiently measured in public research is how the process of finding information translates into workload on production enterprise file infrastructure. Without that connection, we cannot reliably estimate how much infrastructure activity an agent task requires, how it changes with concurrency, or what it costs.
The question is no longer simply whether an agent can find a file. It is what the system has to do while the agent is looking for it.
Sources
1. GitBook (13 May 2026), Research: AI agents are now the majority reader of your docs. Traffic measured 27 April to 3 May 2026.
2. Zhu et al., University of Washington SyFI Lab (25 June 2026), TraceLab: Characterizing Coding Agent Workloads for LLM Serving; see "Autonomous, Multi-Step Conversation" and "Tool-Heavy, Long-Tailed Execution". Paper.
3. Choubey et al., EMNLP 2025, Benchmarking Deep Search over Heterogeneous Enterprise Data.
4. Guo et al., EMNLP 2026, AGORA: An Archive-Grounded Benchmark for Agentic Workplace Document Reasoning.
Methodological note: these sources measure different populations with different methods. No result above should be read as an enterprise file-access cost estimate.
The series
Get the next post.
One email when the next piece in the series goes up. No product news, no sales follow-up.
Unsubscribe in one click. We do not share the list.