# Coding agent

> VI: https://5ac.vn/wiki/coding-agent/ · EN: https://5ac.vn/en/wiki/coding-agent/ · lang: en · spec_as_of 2026-10-19 (spec no formal spec) · cluster C4

AI software that reads a repository, edits files, runs tests and opens pull requests toward a goal — a working entity, not an autocomplete suggestion or a guidelines file.

---
## Definition

A coding agent is AI software that works on code by itself: it reads a repository, plans the change, edits several files, runs tests and opens a pull request without a human issuing each command. Unlike an IDE completion that only suggests text at the cursor, a coding agent takes a goal and finishes the loop on its own.

## Mechanism

The scholarly frame treats the interface between agent and computer as the performance lever: SWE-agent (arXiv:2405.15793) reports that a purpose-built interface lets agents create and edit files, navigate repositories and run tests far better than plain interactive models, reaching 12.5% pass@1 on SWE-bench and 87.7% on HumanEvalFix as stated by its authors. 5ac's on-prem architecture piece splits a coding agent system into four layers, with two warnings attached: without the orchestration layer the system is a chatbot with terminal access, and without the audit layer a company cannot answer the production question of what the agent read, changed and was approved to do.

| Layer | What it does | What to watch |
| --- | --- | --- |
| Orchestration interface | authenticates users, binds role-based permissions, records the original request | a junior engineer does not hold the same task rights as the CTO |
| Execution | reads files, creates patches, runs tests inside a sandbox and a private branch | never edits the main branch or deploys to production |
| Control and audit log | records who asked, which repo was read, which file changed, which command ran | structured entries, so incidents stay traceable |
| Internal data and repos | Git server, package registry, docs, knowledge base | access through a controlled gateway, not personal credentials |

Permissions are granted in four ascending levels: read and explain code, create patches without running risky commands, run tests in a sandbox, then open pull requests that trigger a controlled pipeline. Nobody jumps to level four before logs and rollback exist.

## Application context

Three terms get mixed constantly: AGENTS.md is the markdown file at the repository root declaring build and test commands and conventions for the agent to read before working; Agent Skills (SKILL.md) is the specialist know-how an agent loads on demand; the coding agent is the running entity that reads that file and uses those skills. The lesson for businesses is to pick the species for the job: Codex is the commercial coding agent that delivers now but locks you in, Pi is a minimal harness of roughly 418 lines of agent loop and under 1,000 prompt tokens, and Hermes is the durable assistant that runs a whole company. Same model, different harness, cost per task shifts by more than 2x at equal quality. 5ac's 90-day path measures time from ticket to pull request, share of PRs that clear the first review, errors the agent introduced, and token cost per task; counting lines of code measures the wrong thing.

## Sources fetched

- [On-prem AI coding agent architecture for Vietnamese enterprises](/en/blog/on-prem-ai-coding-agent-vietnam-data-sovereignty/) — four layers, control plane, audit layer — accessed 2026-10-19
- [Product roadmap for the fragmented coding-agent era](/en/blog/product-roadmap-model-agnostic-on-prem-ai-agents-vietnam/) — model-agnostic control layer, four ROI metrics, 90-day map — accessed 2026-10-19
- [Codex, Pi, Hermes — three species of AI agents, not three rivals](/en/blog/three-species-of-ai-agents-codex-pi-hermes/) — three species, Pi's 418-line loop, harness effect on cost — accessed 2026-10-19
- <https://arxiv.org/abs/2405.15793> — SWE-agent: agent-computer interfaces for software engineering tasks — verified via arXiv API 2026-10-19

## Related

- [AI Agent](/en/wiki/ai-agent/)
- [AGENTS.md](/en/wiki/agents-md/)
- [Agent Skills (SKILL.md)](/en/wiki/agent-skills/)
- [Sub-agent](/en/wiki/sub-agent/)
- [Local LLM deployment](/en/wiki/local-llm-deployment/)