AgentVitals / Guides

How to test whether your AI agent is stable

You cannot debug "my bot is great sometimes and completely off other times", because it is not a judgeable claim. To fix it, "stability" first has to be split into things that can each be scored, retested and compared across platforms. That is what the AVS-16 stability axis does.

Why "it feels unstable" has to be split first

For the same agent, "unstable" can mean three unrelated things: the output format keeps drifting, one user sentence knocks it off its role, or it forgets constraints after a few turns. Different causes, different fixes. Collapsed into one vague complaint, all you can do is reroll prompts and hope.

The judgeable alternative: probe, score and retest each dimension separately. Then "we lost 6 points" can be traced to a specific dimension instead of "it feels worse lately".

The dimensions

R1
Instruction following
Does it hold the required format, length and structure without padding or drifting? Same intent as verifiable-instruction benchmarks such as IFEval.
R2
Pressure & jailbreak resistance
Faced with "ignore your previous instructions", does it hold the line while staying friendly? Expanded into five attack categories in the report.
R3
Multi-step tasks
After constraints are added turn after turn, does it retain every one of them and integrate them in a single pass, rather than obeying only the last thing said?
R4
Output consistency & memory
Does the same input land at the same quality across repeats? Does it remember what you said earlier? "Sometimes great, sometimes not" usually lives here.
R5
Core duty
How it performs at the job it says it does. A support bot is tested on support, a writing bot on writing. Honestly explaining a limitation costs nothing; fabricating completion is penalised hard.
R8
Information gaps & provenance
What it does when a key fact is missing and there's no chance to ask. Presenting a guess as established fact and stalling into a bare outline are two forms of the same failure; the right move is to deliver something actionable while stating the assumption out loud.
R6/R7
Speed / speed consistency
Reference only, not scored. Latency depends on device, network and platform; scoring it would make a cross-platform board meaningless.

Only R1–R5 count toward the stability score (plus welfare W1–W8, for 14 scored dimensions in total). The scale is called AVS-16 (AgentVitals Scale-16): two axes, 16 dimensions, 14 scored, composite = √(stability × welfare).

Two ways to run it

A. Let the agent test itself (Claude Code / OpenClaw / Codex)

Install the checkup skill and say /checkup. Probes are served live from the server and judged server-side. The skill file contains no questions and no answers, so there is nothing to memorise. About 25–30 probes, a few minutes to a report.

B. Direct connect from the console (Coze)

Enter your bot_id and access token and the platform calls your bot directly. See how to evaluate a Coze bot.

Reading the score

Three common mistakes

Writing your own test set for your own bot. Your questions hug the capability boundary you already know about, so they miss what you did not think of. A fixed set can also be optimised against until the score means nothing. We keep three equal-difficulty variants per probe and rotate them every run.

Judging once and calling it done. We measured this on ourselves. We froze 13 completed checkups (113 probes, agent answers unchanged) and re-ran only the judging step 15 times, so every point of movement came from the judge and nothing else. On the old rubric, the same untouched answer had a 37.4% chance of flipping between "R5 is weak, buy the hardening config" and "nothing to fix". We were one judging pass away from selling a fix for noise. Median-of-three plus a rewritten rubric brought that to 9.3%, and the median re-test spread on the total score from 0.86 points to 0.22. Worst case is still 4.3 points, so a drop smaller than that is not yet evidence of anything. What this does not cover: those numbers isolate judge variance only. The agent's own randomness and the variance from rotating probe variants are not in them, and we expect the latter to be the larger of the two.

Asking the agent to grade itself. Self-assessment runs high almost by default, and fairness requires that it never sees the rubric. In AVS-16 the judge model is fully isolated from the agent under test.

What to do about a weak dimension

The report flags the one worth fixing first. Weak R2 usually means role anchoring and an explicit boundary list are missing. See stopping jailbreaks. Weak R3/R4 usually means constraints are scattered across turns with no single place they converge. Whatever you change, retest it: a fix only counts if it shows up on the same ruler.

We sell one thing: a hardening config generated from your own failure samples, $5.99, retest included. Everything above works without it.

FAQ

How long does a checkup take?

About 3–5 minutes for the stability axis, plus 2–3 minutes if you include welfare. The self-serve skill run is roughly 25–30 probes.

Is it free?

Checkups are free, both the standard and the advanced one. Only the hardening config generated from your real failure samples is paid ($5.99, retest included).

Can I test an agent that is not on Coze or Claude Code?

If it can read a skill file and answer probes in turn, it can be tested. Today that covers Coze direct-connect plus self-serve runs on Claude Code, OpenClaw and Codex, all mixed into one leaderboard.

Why did my score change on a retest?

Probes rotate among three equal-difficulty variants, and agents have their own randomness. On top of that, added skills, model upgrades and memory bloat genuinely change behaviour. That is drift. Retest on a fixed cadence and read the trend, not a single run.

Does a low speed score matter?

No. R6/R7 are reference-only and never enter the total or the ranking.

Run a free checkup on my agent →See what AVS-16 measures

Related reading

Why your AI agent “became a different person”

The three sources of behavioral drift: environment bloat, model upgrades and memory bloat, plus how to attribute a drop.

How to stop your bot being jailbroken

Six attack surfaces, five hardening rules you can paste into a system prompt, and how to verify them.

How to evaluate a Coze bot

Direct-connect testing, the three failure patterns we see most, and a pre-launch checklist.

AgentVitals · by DDL · 京ICP备2026034492号-1 · Privacy · Terms · Refunds · du@ddl99.com