Spotting drift

Drift is a slow fall in success rate while the suite stays green, or a pass rate that holds while costs climb. Re-run a weekly sample of held-out tasks and plot it against the baseline.

Treat tool, data, and model updates as suspects first. A tendency to blame the prompt wastes weeks when a tool change moved the numbers.

1
2
3
4

baseline success: 0.78; week 4 sample: 0.61
drift = 0.78 - 0.61 = 0.17
review gate (EXAMPLE): drop of 0.10 or more -> review
0.17 >= 0.10, so the gate trips and the week gets a written review

The program below recomputes the drift check from EXAMPLE weekly scores. It uses plain integer points, with no model involved.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

import sys

baseline_pts = 78
weekly_pts = [78, 74, 66, 61]
review_gate_pts = 10

drop_pts = baseline_pts - weekly_pts[-1]
review = drop_pts >= review_gate_pts

if drop_pts != 17:
    print("FAIL: drift arithmetic off")
    sys.exit(1)
if not review:
    print("FAIL: gate should trip")
    sys.exit(1)
print("PASS: drift gate trips on a 17 point drop from 78 to 61")

Reviewing an incident

Every gate trip or bad agent action gets a short written review. The table below fixes the fields so reviews stay comparable.

Least privilege keeps incidents small: narrow file and tool permissions, keep secrets out of prompts and send proprietary RTL only to deployments approved for it, and treat files and logs as untrusted input. Risky actions wait for human approval.

Shutoff rules

Shutoff rules are numeric and owned: when a limit breaks, a named owner switches the agent off. Debate happens when writing the rule, not during the incident.

The calc block below checks EXAMPLE shutoff gates against this topic numbers. Success 0.61 stays on under watch; the other gates stay clear.

1
2
3
4
5

shutoff gate (EXAMPLE): success below 0.60 for two weeks running, or any unapproved write
week 4 success 0.61 is above 0.60, so the agent stays on under watch
one more weekly drop trips the success gate
cost gate (EXAMPLE): cost per resolved task above 0.20 units pauses rollout
0.09 is below 0.20, so the cost gate stays clear

Common mistakes

Related topics

• Agent regression suites

• Data, privacy and IP