Agent Skills and Grounded Context

An agent skill can teach a model how to perform a task. It cannot, by itself, tell the model what is true about your current software.

Workspai keeps those responsibilities separate:

• AGENTS.md and generated skills explain how an agent should work.
• The workspace model defines which projects and boundaries belong to the system.
• The graph lets the agent retrieve focused relationships and supporting proofs.
• Doctor and verification artifacts show what is healthy, blocked, stale, or still unknown.

This means every supported agent can start from the same workspace understanding without receiving the entire repository in every prompt.

Prepare the workspace for an agent

Run the intelligence loop from the workspace root:

npx workspai workspace intelligence run --for-agent generic --strict --json

Workspai refreshes the governed artifacts and writes the agent-facing entry points. If you only need to refresh those surfaces, run:

npx workspai workspace agent-sync --write --refresh-context --preset enterprise

The generated files point the agent toward bounded evidence. The agent can then query the graph itself when a task needs more detail; a user does not need to paste the full graph or manually run every search command.

What the agent receives

The exact files depend on the workspace contract, but the handoff includes:

• current workspace and project boundaries;
• approved commands and repair capabilities;
• focused context and skill indexes;
• links to model, graph, Doctor, verification, and explanation evidence;
• freshness and unknown-state signals that prevent stale context from looking complete.

What Workspai does not promise

Generated instructions do not guarantee that every model will obey every step. Workspai makes the task inspectable: the workspace owns the truth, the agent retrieves bounded evidence, and verification decides whether the result is acceptable.

Continue with Workspace Doctor to see how blockers become repair evidence, or Verified AI Repair to follow a repair from intent to proof.