Is git diff still the right thing for humans to review when coding agents work faster than we can?

I've been thinking about a problem that seems to become more important as coding agents get faster and more autonomous. Traditional code review assumes that the main thing a developer changes is the code. Someone writes code, another person reads the diff, tests run, and we decide whether to merge. But an agent can do much more than produce a diff. During one task it might: run dozens of shell commands install or remove dependencies start background processes access the network modify local configuration run migrations interact with Git/GitHub spawn subagents fail, change its plan, and retry touch things that don't appear in the final diff at all At that point, the Git diff is only one projection of what the agent actually did. The obvious answer is "look at the transcript / tool logs", but I don't think that scales either. If an agent makes 100 tool calls and changes 20 files in a few minutes, reading the entire execution trace defeats much of the point of having the agent. This makes me wonder whether the human-facing side of an agent harness is still underdeveloped. A harness today is usually discussed in terms of what it does for the agent: tools, permissions, context, sandboxes, subagents, memory, retries, verification, etc. But maybe a mature harness also needs to do something for the human supervising it: turn a large amount of agent activity into something that a human can actually reason about. For example, after a task I might care more about questions like: What was the agent allowed to do? What did it actually do? What persistent state did it change? Did it perform any external or high-consequence actions? What verification actually ran? Where did it deviate from its original plan? What can the system prove, and what is still uncertain? Can I drill down into the exact diff / command / trace when something looks suspicious? Something like: The important part would be that these facts come from runtime/tool evidence, not just from the agent saying "here's what I did." I'm not proposing a specific product or dashboard yet. I'm mostly wondering whether the unit of human review needs to move upward as agent throughput increases. Instead of reviewing every line and every action, maybe humans increasingly review the agent's consequences, decisions, evidence, and exceptions, with code and raw traces available as drill-down. For people already running coding agents heavily or building your own harnesses: What do you actually want to inspect before you trust an agent's work? Do you still rely mostly on Git diff + tests? Do you inspect tool history? Have you built summaries, checkpoints, sandboxes, audit logs, approval gates, or something else? And where do you think the line is between useful compression and a misleading "everything is fine" dashboard?