Research paper 15
Proof-Carrying Development: Agentic Code Work as Claims With Attached Evidence
AI coding agents do not merely write code. They make claims about what changed, what was inspected, what proof ran, what remains risky, and whether a task is complete. Those claims are often more important than the generated code because they shape user trust and downstream action. This paper proposes proof-carrying development as a pattern for agentic software work. In proof-carrying development, every nontrivial closeout binds the change to source state, touched surfaces, proof commands, observed results, claim ceilings, and remaining risks. The local evidence comes from the aggregate-only A Longitudinal Corpus of Human-Codex Software Work descriptor by A.G. Mauro and C.A. Harris and portfolio workflow artifacts. A refreshed aggregate at snapshot cutoff 2026-07-21T07:19:35.322Z found 7,514 session files, 7,183 sessions with token information, 331 without token information, 174,068 token-count events, 2,060 null token-count events, 245,480 function calls, and 238,745 shell command calls. The logs show that tool-mediated development produces abundant evidence, but that evidence only helps when closeouts name what it proves. The paper contributes a claim-and-proof schema for agentic development: treat the final answer as a receipt, not a victory lap.
- Paper
- 15
- Authors
- A.G. Mauro and C.A. Harris
- Date
- 2026-07-21
- Collection
- Standing Framework Research
Abstract
AI coding agents do not merely write code. They make claims about what changed, what was inspected, what proof ran, what remains risky, and whether a task is complete. Those claims are often more important than the generated code because they shape user trust and downstream action. This paper proposes proof-carrying development as a pattern for agentic software work. In proof-carrying development, every nontrivial closeout binds the change to source state, touched surfaces, proof commands, observed results, claim ceilings, and remaining risks. The local evidence comes from the aggregate-only A Longitudinal Corpus of Human-Codex Software Work descriptor by A.G. Mauro and C.A. Harris and portfolio workflow artifacts. A refreshed aggregate at snapshot cutoff 2026-07-21T07:19:35.322Z found 7,514 session files, 7,183 sessions with token information, 331 without token information, 174,068 token-count events, 2,060 null token-count events, 245,480 function calls, and 238,745 shell command calls. The logs show that tool-mediated development produces abundant evidence, but that evidence only helps when closeouts name what it proves. The paper contributes a claim-and-proof schema for agentic development: treat the final answer as a receipt, not a victory lap.
← Back to research papers