CLiF turns product usage into evals

CLiF targets the biggest weakness in agent development: teams rarely have enough realistic data to evaluate how agents perform inside real workflows. Turning product telemetry into a learning loop is compelling, but privacy, consent, and noisy behavioral signals will determine whether this becomes infrastructure or just another observability dashboard.

• –Full-stack trajectories capture both agent decisions and the user context surrounding them.
• –Production-derived evals should expose failure modes synthetic benchmarks routinely miss.
• –Approved skills can update agents without redeploying, shortening the iteration cycle.
• –Self-hosting and exportable artifacts make the approach more credible for enterprise teams.
• –The key risk is confusing frequent user workarounds with genuinely reusable agent behavior.