Update: I'm exploring the idea of what I'm calling an "autonomous bot" for my farm.
From my research, it's clear a key component of BotWatch is sample size. This is to say that if you ran a script for, say, 2 hours a day, the ban rate is low. Why? Sample size. Jagex needs a large data sample from your account to issue a ban with confidence, and confidence is key. As a games company, they can't afford to incorrectly ban a large number of players for botting. This is likely why limiting your daily botting time helps avoid bans, and why suicide botting sessions reliably produce bans. In the first case, they have little data to analyze, and in the latter, they have more than enough.
This is, of course, a hunch, but one I'm confident enough to make a new Script around. An autonomous, AIO bot that does the following:
1. Has a list of Tasks that can be performed, with requirements for each task, setup for each task (getting items from bank, traveling to area, etc), and task completion (teleporting away from an area, depositing outputs into a bank, etc)
2. Randomly chooses a task to execute based on whether or not requirements are met, or whether or not the the task is still appropriate given the account's current stats (perhaps the account is too high of a level for the task in question to be worth doing), etc.
3. Progresses an account in a very generalized sense, choosing some activity or skill, executing it for 1 - 2 hours, and then changing tasks.
4. Does herb runs, if able
An additional benefit of this approach is that it further mimics human behavior via task switching. Humans aren't going to sit and do the same activity for 10 hours straight without breaks, and neither will this bot. Prolonged sessions of the same activity are a dead giveaway that an account is a bot. Autonomous, randomized task switching is an approach I've not seen tried anywhere else, and so I'm willing to give it a go.
This is obviously an ambitious project, and will rely heavily on code reuse and object oriented programming to be doable.
TL;DR: I'm looking into making a bot that mimics human behavior in a very macro sense. It'll train a variety of skills and use some sort of simple "AI" to decide what tasks it should be doing, switch tasks randomly, and not focus on a single activity. Most activities in the activity list will be profitable, but not all.
PS: For the people wondering, I've not given up on the project yet, I'm just trying to find an approach that's more viable than suicide botting. For right now, that means a lot of theorizing about how detection works, testing my hypothesis, and doing research into other people's "botting to max" series, since they emphasize the longevity of an account.