Language

$AXAgent Experience

The discipline of measuring, shaping, and improving what an AI coding agent perceives, remembers, and acts on across sessions. The direct analog of Developer Experience (DX) for human engineers.

in the wildAX is to AI agents what DX is to humans.

$AXLAgent Experience Layer

A software layer that owns the AX surface - ingest, storage, query, and agent-readable interfaces over the evidence of prior agent behavior. ax is the reference implementation.

in the wildAn Agent Experience Layer needs typed evidence, multi-agent ingest, and an agent-readable interface.

$ax

The project. Lowercase, always. The whole thing - the agent experience layer over your coding-agent history.

in the wildax watches your sessions and proposes small, repo-specific fixes.

$axctl

The CLI binary under the hood. You mostly type ax; axctl is the technical name of the executable.

in the wildThe ax command you run is the axctl binary on disk.

$ax studio

The live local dashboard. The improve-first view of your graph: what's next, the proposal deck, experiments, cost and routing.

in the wildOpen ax studio to triage proposals and watch experiments land.

$retro

A structured reflection collected after a session: what was tried, what worked, what failed, and the next experiment to run. It is a bet on the next session, not a recap. ax drains pending retros in the background (watcher-driven) and through the /retro skill - never via a per-turn Stop hook.

in the wildSix retros this week propose the same hook - that's a pattern worth shipping.

$proposal

A repeated-mistake pattern ax mines from your history, surfaced as one small candidate fix - a skill, a hook, a piece of guidance, or a routing change. Reviewed one at a time, never auto-applied.

in the wildax found a 26x recurring mistake and turned it into one proposal.

$experiment

What an accepted proposal becomes: the scaffolded artifact plus a checkpoint schedule. ax measures whether it actually helped at +3, +10, and +30 sessions.

in the wildPast bets, measured - this experiment is +10 sessions in and holding.

$verdict

The outcome locked at a checkpoint. One of: adopted (doing real work), ignored (created but never invoked), regressed (made things worse), partial (mixed signal), or no_longer_needed (the pattern self-resolved).

in the wildI locked the schema-guardrail verdict as adopted.

$ax-loop

The closed self-improvement cycle: retro -> proposal -> experiment -> verdict -> the next session reads what worked. "Close the loop" means one full pass from retro to verdict.

in the wildReceipts over vibes - the ax-loop only ships fixes it can measure.

$ax-graph

The local typed evidence store behind everything: sessions, turns, tool calls, skills, commits, retros, proposals, experiments, verdicts. The abstraction is "the ax-graph"; the database is an implementation detail.

in the wildEverything ax shows you is read back from the local ax-graph.

$impact / projected value

The backtested estimate of what a proposal is worth, computed from your own history - e.g. "$605 redirectable" or "26x recurring." ax reprices what already happened from real tokens; it never reports fabricated savings.

in the wildThis proposal's projected value is $605 redirectable spend.

$dispatch

A sub-task your agent spawns - the routine work that runs on whatever model the harness picks. Left alone, dispatches default to your most expensive model.

in the wildax dispatches shows which sub-tasks ran on the expensive default.

$routing class

A named bucket of look-alike dispatches (file search, well-specified implementation, bug fixes) that can safely tier down to a cheaper model. Judgment work - review, design, planning, audits - never routes automatically.

in the wildax routing tune mined a new routing class from work you keep overpaying for.

$quota

Your live Claude plan usage across the 5-hour and 7-day windows. ax surfaces it in the CLI, your statusline, and the menubar so you see the limit before you hit it.

in the wildax quota --statusline puts your plan usage one glance away.

$recall

Full-text search across your turns, commits, and skills. Auto-scopes to the current repo when run inside a git tree. The fastest way to ask "have I solved this before?"

in the wildax recall "surreal index drift" --scope=here pulled the fix from a session last month.

$hook

A typed guard authored once in Effect TypeScript and run on both Claude Code and Codex via the ax hooks SDK. Verdicts are allow / block / warn / inject; a defect fails open. This is how one person's trick becomes a deterministic team-wide rule.

in the wildThe route-dispatch hook warns the moment a routine sub-task is about to run expensive.

$ax skills weighted

The ranking of your skills by real usage times their role weight - the honest answer to "which skills actually earn their place?" Replaces any older "taste score" phrasing.

in the wildax skills weighted surfaced three skills I install everywhere but never run.

$profile

A shareable summary of one seat - stats, rig, and taste patterns - published with ax profile publish. The consent prompt shows you the exact JSON before anything leaves your machine.

in the wildax profile publish posted my profile to a gist after I approved the payload.

$leaders

The public community boards compiled from registered profiles - leaderboards and trending skills at /leaders, with each member's page at /u/<login>. Aggregates only; no code, no transcripts.

in the wild/leaders shows trending skills across everyone who opted in.