Research paper 17
Tool Calls Are the Medium: Agentic Coding as Human-Agent-Tool Choreography
Agentic coding is often described as conversation: a human asks, a model answers, and code appears. A Longitudinal Corpus of Human-Codex Software Work by A.G. Mauro and C.A. Harris suggests a different object of study. At snapshot cutoff 2026-07-21T07:19:35.322Z, the local corpus contained 7,514 session files, 22,923 user messages, 79,110 assistant messages, 245,480 function calls, and 238,745 shell command calls. The top aggregate function families were exec_command at 203,133 calls, write_stdin at 35,612, and update_plan at 3,889. This paper argues that tool calls are not incidental implementation details. They are the medium through which agentic development becomes inspectable: file search, source reading, patching, validation, git-state inspection, runtime probing, proof reruns, blocked-state preservation, and closeout receipts. The contribution is a measurement frame for agentic software work that treats human message, assistant response, tool family, source state, and proof outcome as one choreography rather than separate artifacts.
- Paper
- 17
- Authors
- A.G. Mauro and C.A. Harris
- Date
- 2026-07-21
- Collection
- Standing Framework Research
Abstract
Agentic coding is often described as conversation: a human asks, a model answers, and code appears. A Longitudinal Corpus of Human-Codex Software Work by A.G. Mauro and C.A. Harris suggests a different object of study. At snapshot cutoff 2026-07-21T07:19:35.322Z, the local corpus contained 7,514 session files, 22,923 user messages, 79,110 assistant messages, 245,480 function calls, and 238,745 shell command calls. The top aggregate function families were exec_command at 203,133 calls, write_stdin at 35,612, and update_plan at 3,889. This paper argues that tool calls are not incidental implementation details. They are the medium through which agentic development becomes inspectable: file search, source reading, patching, validation, git-state inspection, runtime probing, proof reruns, blocked-state preservation, and closeout receipts. The contribution is a measurement frame for agentic software work that treats human message, assistant response, tool family, source state, and proof outcome as one choreography rather than separate artifacts.
← Back to research papers