Agent observability, in three layers.

Trace every agent run end to end: tools, prompts, inputs, and outputs. Replay any execution and follow a failure back to the step that caused it.

Ask in plain language why a run broke, in the console, in Slack, or from the SDK. Cortex reads the trace and cites the exact tool call, prompt, and context behind every answer.

The behavioral record of every run: events, evidence, actors, outcomes, and repeated patterns become typed memory that Eyes writes and Cortex reads.

Register internal HTTP APIs once, then query them in plain English from the terminal, by humans or AI agents, with a deterministic regex fast-path. Install with npm i -g @invariance/q.

tars hooks into Claude Code, Codex, and Cursor, mirrors the structure of your cleanest sessions, and can feed self-relative coaching back into the agent. Anonymous, opt-in, local-first.

gps gives Claude Code, Codex, and Cursor durable repo memory: notes, decisions, preferences, invariants, and tests anchored to the symbols an agent is about to touch.