KN-05 · The harness tax

What you pay before your first word

A Claude Code session can send a system prompt and tool definitions before you type the task. The study measured that startup overhead, then added tools back to see what they contributed.

November 6, 20265 min readGuildBuild Team

AI agentsHarness engineeringToken measurementAgent economics

In this article

What a bare seat sends

On one Windows 10 machine, the team removed its instruction file, denied built-in tools, disabled skills and memory, and turned off the optional features it found. That left a baseline for each surface.

First call from a bare seat
Surface and release	Input tokens
Terminal · Claude Code 2.1.282	1,884
Terminal · Claude Code 2.1.283	1,887
Desktop app · 2.1.282	8,633

The desktop baseline was more than four times the terminal baseline. About 58% came from the app’s own servers. The team did not find a setting that removed those servers.

Figure 01Field measurement

Desktop sent more input before the task

First call

input tokens

Terminal 2.1.2821,884

Terminal 2.1.2831,887

Desktop 2.1.2828,633

One stripped-down Windows 10 setup in late September 2026. About 58% of the desktop baseline came from the app’s own servers.

Field note evidence and limits

What the tool definitions add

The team enabled built-in tools individually in the terminal and used the harness’s context report to estimate what each added to startup.

Estimated startup tokens per tool
Tool	Tokens
PowerShell	3,369
Grep	1,269
Bash	1,069
Read	569
Edit	369
Glob	269
Write	269

PowerShell alone added more than Bash, Grep and Read combined. In the tested setup, PowerShell came enabled by default on Windows. Denying a tool removed its startup cost in this test. Of the 44 built-in tools considered, 36 added zero in this setup because they were deferred, already denied or absent from the terminal.

Figure 02Field measurement

PowerShell was the largest tested tool definition

Individual tools

estimated startup tokens

PowerShell3,369

Grep1,269

Bash1,069

Edit369

Glob269

Write269

Estimates from the context report on one Windows terminal setup, accurate to about 50 tokens in this study. Deferred, already denied or absent tools added zero.

Field note evidence and limits

Startup totals can shift together

Five worker seats started at roughly 56,000 to 67,000 tokens before one change and roughly 20,000 to 25,000 afterward. The lead seat shifted too.

Startup input before and after the shift
Seat	Before	After
Worker 1	66,130	24,209
Worker 2	67,427	23,484
Worker 3	57,562	23,769
Worker 4	56,203	23,149
Worker 5	58,447	21,221
Lead seat	60,106	39,815

Different instruction files showed a shift on the same day, which points to a possible shared cause. The harness release, local settings and model all changed together. The study could not isolate which change mattered.

Your shared instructions add a second layer

An earlier GuildBuild reference fleet loaded a 58 KB instruction file, about 47,000 tokens, on every seat. Agents could still work. The operator judged that costs stalled the fleet, based on the bill and their experience. This was an uncontrolled observation.

If a resident block appears in every call, estimate the volume it adds by multiplying its tokens by the number of calls across seats. Convert that volume to cost using actual cache reads, writes and ordinary input rates. A shared instruction file’s impact can grow with both seat count and session length.

Set a baseline you can remeasure

Measure a stripped-down first call on the surface you actually use.
At the same release, subtract that floor from the configured seat’s first-call total.
Check whether disabling unused tools changes the measured total. Test large definitions first.
Measure again when the harness, model or settings change. Estimate how widely a resident instruction block will be reread.

How we know

The measurements came from one Windows 10 machine in late September 2026. The terminal floor was measured at two versions. Individual tool values came from the context report and were accurate to about 50 tokens per the study. The before-and-after table drew on 117 sessions. Its cause is unresolved. We did not rerun the original measurements for this publication.