AI coding agents

An agent in Jarvis is a teammate on the bench: a name, a role, a set of instructions that say how it works, and an AI model. When you assign a ticket to an agent, its settings decide how it approaches the work and what it is told before it reads the ticket.

This page covers the bench, how to choose an agent, and the two ways to teach agents: knowledge and skills.

#Claude Code and Codex

Agents write code with one of two AI coding tools: Claude Code or Codex. The agent's model decides which. Both work on the ticket's own branch, both stop at the approval gate, and both show their progress on the same run screen. From your side, the only difference is the agent you pick.

If an agent's model cannot write code, or no coding tool is available to it, the run writes a plan instead of changing files. That is still useful for scoping a ticket.

#The bench

The Agents screen lists every agent. The workspace comes with the roles AiM Infosoft staffs: Laravel, PHP, frontend, database, API integration, QA, DevOps, code review, Angular, .NET, React, Node.js, Python, mobile, Flutter, React Native, Java, WordPress, Vue, HTML, UI design, security, performance, SEO and support engineers, plus a tech spec writer. Three agents use Codex: Codex Engineer, Codex Refactorer and Codex Test Engineer.

They all build on Baseline Standards, a shared set of rules about quality and honesty that every agent follows.

Your administrator can add a larger library of personas from the community. They arrive switched off, so the assign dialog stays short, and can be switched on one at a time from the Agents screen.

#What an agent has

Field	What it does
Name and Role	How people recognise it on the assign dialog.
Division	A grouping, used to filter a large bench.
Inherits from	Another agent whose rules are added to this one's.
Model	Which AI model it uses, and so which coding tool.
Effort	How hard the model thinks, from low to max. More effort costs more.
Instructions	The agent's persona and rules, written in the editor.
Status	Only active agents can be assigned.

Every project has a default agent. A ticket gets that agent unless someone picks another.

#Choosing an agent for a ticket

• Pick the specialist for the main change. A ticket that is mostly database work goes to the Database Engineer, even if it touches one page.
• Plan before you build when unsure. Use Analyse only or Prepare implementation plan for a ticket you are not sure how to scope. A plan costs less than a wrong implementation.
• Compare. Try a Claude Code agent and a Codex agent on similar tickets and compare their changes. The gate makes the comparison safe.

#Knowledge

Each agent has a knowledge base: short documents it should know, such as a coding standard or a guide to a framework. An entry can apply everywhere or to one project.

When an agent picks up a ticket, the entries most relevant to that ticket are given to it. Relevance is judged mostly by the entry's title, so give entries clear, specific titles.

Use knowledge for what belongs to the agent. Use notes for what belongs to the project.

#Skills

A skill is a written procedure an agent can follow: a checklist for a security review, a recipe for adding a new page, the steps for building from a design. Skills live in the Skills library.

• Draft and published. A skill starts as a draft. Only published skills are given to agents.
• On an agent or a project. A skill on an agent goes with it to every ticket. A skill on a project is given to every agent working on that project.
• Shared, and forkable. Everyone can read the published library. To change someone else's skill, fork it and edit your own copy.
• Supporting files. A skill can carry templates and reference files alongside its instructions.

The library can also import a set of open-source engineering skills in one go.

Some people work only on skills. Their group sees the skills library and nothing else, on a site of its own.

#When an agent cannot start

If a run says the agent could not sign in, or that no coding tool is available, there is nothing to fix on your side. Tell your administrator; the run screen will offer Retry once it is sorted.

Next: What an agent reads.