ORBiS Tree

ORBiS TREE · ENGLISH DOCUMENT

Concepts →

Agentic Coding

AI software development that joins repository inspection, edits, execution feedback and verifiable completion criteria.

한국어 English
English updated 10월 18, 2026 Source article updated 10월 18, 2026 2 sources
This English page is a curated translation layer linked to the Korean source article. Community changes are currently made on the Korean source, where the full revision history and anonymous edit trail are preserved.

Overview

Agentic coding extends code suggestions into repository inspection, edits, tool execution and feedback. Distinguish a product such as Codex from the general design of an AI Agent. Anthropic's agent design article ↗ discusses test feedback and review in coding.[2] It is a 2024 design account, not a current menu or performance table.

Turn a task into a reproducible change

ORBiS Tree proposes recording reproduction inputs, expected outcomes, repository version and environment. Locate the failure, change the necessary files and run relevant checks. Connect the diff, commands, results and remaining failures. When public deployment is requested, verify the deployed result separately.

Fictional case: preserve monetary input conditions

A fictional integer-won parser should map '12,000원' to 12000 and '0원' to 0, while rejecting '금액 미정'. Removing every non-digit could incorrectly accept '미정12원'. Define currency, grouping and negative-value rules before checking both accepted and rejected inputs.
These examples do not define a real service's currency policy. Tests copied from an implementation can reproduce its mistake. Derive expected outputs independently from requirements.

Worked example: include regressions in success

For twenty fictional issues, fourteen fixes satisfy requirements and regression checks, three fix the target but break another behavior, and three remain incomplete. Overall success is 14/20=70%. Successful shell commands or generated lines do not replace the task denominator. This is not a measured performance result.

Benchmark scope and product evaluation

The SWE-bench paper ↗ evaluates repository changes against GitHub issues.[1] Appendix A.4 checks both failures that should become passes and existing passes that must remain passes. Its original repositories, language and execution conditions differ from a whole production system. A benchmark score does not establish deployment quality.

Repository material and execution authority

Project guidance can inform a task without making every file authoritative permission. Consider Prompt Injection in external issues or dependency material, and enforce the actual execution scope through AI Agent Oversight and the environment. Retain separate evidence that code was saved, checked and published.

SOURCES

Sources

  1. Open source ↗
  2. Open source ↗