· via dev.to (home feed)
Atlarix 15.18: agent writes its own tools, work modes trimmed from five to three
Norah Labs' AI coding agent Atlarix 15.18 can now author its own panels, commands and rules, while Debug and Review modes become /debug and /review commands usable in any mode.

Atlarix 15.18 arrived on November 15 with two changes that reshape how the AI coding agent is used day to day: the agent can now build its own tools, and its work modes have been consolidated from five down to three.
Atlarix, built by Norah Labs, plans, edits, executes and reviews code inside a user's own projects, and it requests confirmation before it modifies any file. According to the release announcement published on dev.to, this version is about lowering friction: fewer questions about routine actions, agent-generated tooling, and easier ways to take your work and data with you.
The agent writes its own tools
The headline feature lets users ask Atlarix for a panel, a slash command or a rule, and have the agent produce it. These are saved as either a silk, a reusable file combining instructions and commands, or an edict, a rule governing what the agent is permitted to do. Every generated file still requires approval before it lands in the project, presented as one confirmation card per file, so users can see exactly what is being added.
Mods also received an upgrade. A mod, defined as a small program that adds a panel or command to Atlarix, now runs as real code in its own process through the product's mod API, which means its panels update live rather than sitting static.
Five modes become three
Until this release, Atlarix asked users to choose among Explore, Plan, Build, Debug and Review before typing anything. Only the first three remain, while Debug and Review have been converted into /debug and /review commands that can be invoked from any mode.
A companion dev.to post from the Atlarix team explains the reasoning with usage data. Across three weeks of real turns from October 26 to November 14, excluding the company's own and test accounts, roughly 84% of turns ran in Build, 11% in Explore and 4% in Plan. Debug was used for a single turn; Review for none. A choice people skip still costs them something, the team argues: one more thing to read and one more question before they can start.
Each remaining mode now has a distinct job. Explore answers questions about the code and reads files without changing anything. Plan works out the steps for a change and presents them for review, also without touching the code. Build is the only mode that edits code, which makes the current mode an at-a-glance indicator of whether Atlarix can modify the project. Requesting a change from Explore or Plan prompts Atlarix to ask before it switches to Build. Shift+Tab cycles through the three modes in the app and CLI, and the same trio appears in the VS Code, Zed and JetBrains integrations. The team also notes that Codex and Claude Code similarly treat review as a command rather than a mode.
Backward compatibility is handled quietly: chats previously saved in Debug or Review now open in Build, as do any settings that still reference a removed mode. Nothing saved is lost, and no migration is required.
Fewer interruptions and easier exits
Permission handling is simplified in the same spirit. Everyday commands such as ls now run without asking, while anything that alters the machine still requires confirmation first.
New commands round out the release. /export saves any chat as Markdown from the app, the CLI and VS Code, and /bug files a problem report from the point where the issue occurred, with the chat attached.
Account deletion is available from the profile menu in the app, via atlarix account delete in the CLI, or from the account page on the company's site. It requires typing an email address to confirm, and re-authenticating if the session is older than ten minutes. Deletion is immediate, signs the user out everywhere, cannot be undone and forfeits any remaining credit on the account.
On pricing, Atlarix Auto remains free with a daily limit and now tells users when that limit resets. Pro subscribers get $19 of Atlarix Core usage each month, charged at the same rates as credit, with pay-as-you-go billing beyond that.
Why it matters
Two currents in AI coding tooling meet in this release. First, agent-authored tooling shifts configuration toward conversation: instead of hand-writing extension files, users describe what they want and the agent generates it, with approval kept in the loop. Second, the mode consolidation, justified with usage data and mirrored by Codex and Claude Code, suggests the field is converging on a small set of states plus on-demand commands rather than sprawling modes. Combined with a clearer permission line and stronger data portability, 15.18 is less about raw new capability and more about an agent that interrupts its user less, which for people coding with these tools daily may matter just as much.
- #ai-coding-agents
- #developer-tools
- #release-notes
- #cli