Context artifacts & grounding

The most expensive part of dispatching an agent is not the work — it is the warm-up. A fresh session knows nothing about your task: not the goal, not the relevant files, not the constraints you settled three sessions ago. Left alone, it spends its first stretch re-discovering all of that, and re-discovery is where drift and wasted spend live.

Context artifacts remove the warm-up. An artifact is a curated context package you build once and attach to any number of sessions — so an agent's very first prompt already lands on a briefed mind.

What an artifact is

A context artifact is a Markdown document with three kinds of content:

• Curated notes — the task framing, decisions already made, constraints, vocabulary. Whatever you would otherwise re-explain.
• File references — pointers to the documents and code that matter, so the session knows exactly where to look instead of searching.
• An optional attached knowledge graph — a purpose-built Cortex workspace bundled with the artifact. When you build one, you pick the graph depth: structure only, structure plus entities, or full.

Artifacts are authored in the Context Artifacts view, and they live as ordinary files in your workspace — which means agents can also read and even maintain them, and your edits show up in the view automatically.

Grounded sessions

Attaching an artifact to a terminal session produces a grounded session: a full, persistent, interactive Claude Code session whose system context includes your artifact from the moment it starts.

The boundary is precise and worth understanding: only the context is injected. A grounded session keeps its complete capabilities — all its tools, its own model settings, its own permissions. Grounding changes what the session knows at start, never what it can do. That is what makes grounded sessions the workhorse of serious delegation: full power, zero warm-up.

Grounding is captured when the session launches. Editing an artifact later does not reach into sessions already running — start a fresh session (or respawn the pane) when you want it to carry the updated context. Deliberate, and the right default: a live agent's briefing never shifts under its feet mid-task.

Attach where the work happens

Context attaches at two levels in the Terminals view:

• Workspace-level — the default context for every agent pane in a terminal workspace. Right for a fleet: attach the mission context once, and every worker inherits it.
• Pane-level — context for one specific session, layered on top of the workspace default. Right for a specialist that needs its own briefing.

A session's effective context is the combination of both.

One mechanism, every door

The same grounding machinery serves every consumer: sessions you launch from the interface, sessions Voice stands up for you, and sessions agents spawn for other agents all resolve artifacts through one backend. An agent-dispatched worker inherits the workspace's attached context exactly the way an operator-launched one does — by construction, not by convention.

This is what makes artifacts a genuine operator tool rather than a convenience feature: they are how your framing of the work travels into every session that touches it, no matter who or what opened the session.

The habit that pays

For any effort that will outlive one session, spend the first minutes building its artifact: the goal, the decisions, the file pointers. Attach it to the workspace where the work happens. From then on, every session you or the system opens against that effort starts briefed — and when the framing changes, you update one document instead of re-explaining yourself.

Where to go next

• The knowledge graph (Cortex) — building the graphs artifacts can carry
• Missions, operations & the fleet — where grounded sessions do their biggest work
• The operator model — why context packaging is the operator's highest-leverage move