Agent Intelligence is empty and nothing is captured

Learning is a chain, and with any link missing the screens look the same. Waiting is not the cure. The links:

1. Learning is switched on for the scope the installation covers. An installation reads that scope and does not look below it, so a switch set on one repository does nothing for an installation that covers the organization above it.
2. The installation was written after the switch. Switching learning on writes nothing by itself. The hook reaches the client only when the installation is added or updated.
3. The hook is in the client settings of the folder the installation writes into, which is often the folder above the repository.
4. The hook can send events. If the address changed, or another program took it, the events go nowhere and the agent is not told.
5. The installation was written before its hook carried the intelligence. Events arrive and are discarded.
6. Naming is manual by default, so events are queued and nothing is named.

Only Claude Code, Codex and OpenCode take the learning hook.

1. On Harnesses Manage, read the row of the installation. It says which link is broken, for example “not capturing”, “nothing installs this scope”, “no hook in the harness settings” or “hook written, nothing has arrived here yet”.
2. Switch learning on at the scope the installation covers. Then open the installation, click Edit and Save, or click Update installed skills. Nothing else writes the hook.
3. If the badge names the date of the last event that was kept and tells you to update the installation, update it. Updating rewrites the address and the hook even when no file changed.
4. If events are queued and nothing is named, click Look for patterns. Raw events are dropped after a retention period, so a long gap loses material.
5. Work in the folder of the installation. A row that says “hook firing” with the counts of arrived and kept events is working.