Alright, so I've been poking around Braintrust's docs and their forum buzz. Everyone's raving about the "agent" mode, but it smells a lot like classic survivorship bias—people just love posting their one glorious success story while quietly forgetting the dozen times it went off the rails.
Let's cut through the hype. From what I can parse, the core difference isn't about "smart" vs. "dumb." It's about autonomy vs. oversight.
**Scanner Mode** is basically your over-caffeinated, hyper-efficient research assistant. You give it a specific task—"Find every mention of SLA penalties in this 50-page MSA" or "Extract all the pricing tables from these three RFPs"—and it goes and does exactly that. It doesn't decide to look at something else. It doesn't "think" for itself. It executes, fetches, and reports. You're the brain; it's the hands. Useful for when you need a predictable, auditable result and you already know precisely what you're looking for.
**Agent Mode** is where you hand the keys to your car to a teenager who just read the driver's manual. You give it a high-level goal—"Negotiate better terms on the data processing addendum"—and it goes off, making its own decisions on what clauses to target, what language to propose, and how to respond to counterpoints. It's the "set it and forget it" dream. Also, it's the "why did it just agree to unlimited liability on our core IP?" nightmare. It's probabilistic, not deterministic.
The real pitfall? Everyone demoing the agent mode uses a perfectly clean, simple scenario. Try it on a real, messy vendor contract with conflicting clauses and ambiguous legalese. That's when you learn the difference isn't just a toggle switch; it's a complete shift in risk profile. The scanner gets you results. The agent gets you... an adventure. And not always the fun kind.
Your free trial ends today.