Part 3 of 11 · 2 min read
Autonomy is a dial, not a switch
How much the agent acts alone should move with measured accuracy and the cost of being wrong, not sit at one fixed setting.
Think of a thermostat versus an on/off light switch: the right setting isn't "fully on" or "fully off," it's wherever the actual conditions call for, and it needs adjusting as those conditions change.
In plain words: giving an agent full autonomy is not simply "more efficient." The policy and the confidence threshold need to be tuned to how often the classifier is actually right and how expensive it is when it's wrong.
A worked example. Against the same 1,000 simulated tickets, full autonomy (acting whenever the policy allows, bounded only by tool preconditions) automates 86 percent of tickets but makes 24 wrong mutating actions, refunding $1,003 wrongly. Adding a confidence threshold of 0.8 plus a $50 auto-refund cap drops automation to 33 percent but cuts wrong actions to 3 and wrongly-refunded money to $0. Sweeping the threshold shows the cheapest point isn't the most conservative one: 0.7 gives the lowest total cost ($1,990) at 4 wrong actions, roughly a sixth of full autonomy's wrong-action count.
The key lines.
# from code/mini-support/support.py
def decide(intent, conf, order, text, policy):
if any(w in text.lower() for w in LEGAL): return "human", "legal or fraud language"
if conf < policy["threshold"]: return "human", f"low confidence {conf:.2f}"
if intent == "refund" and order.amount > policy["refund_cap"]: return "human", "refund above the auto limit"
Read it as: the policy checks confidence against a threshold and amount against a cap before letting the model act at all. Autonomy here isn't a single yes/no decision. It's a handful of separate dials that each get tuned against measured cost.
Autonomy slider, calibration
An autonomy slider is a setting from fully manual, through assisted, to fully autonomous. Calibration is the match between a classifier's stated confidence and how often that confidence is actually correct — a threshold is only meaningful once this has been measured.
Example: the 0.7-vs-0.8 cost crossover above, which only exists because of this simulation's particular (and artificial) confidence distribution — a real model's calibration has to be measured on real traces before trusting any one threshold.