---
title: "Which Agent Number to Move First"
description: "Weak public score, lost recommendation, and failed integration each call a different first fix among readiness, preference, and task completion."
canonical: "https://www.agentrail.co/blog/agent-experience-metrics"
author: "Sidhant Goyal"
published: "2026-11-03T00:00:00Z"
last_updated: "2026-11-03T11:27:13.732Z"
---

# Which Agent Number to Move First

Weak public score, lost recommendation, and failed integration each call a different first fix among readiness, preference, and task completion.

![Five metric tiles on a blue gradient, with 48 percent lifted above 62 percent, 1.8 seconds, 3.4 times, and 91 percent](https://images.ctfassets.net/1htfyfnn1o5y/2D8hSlC2Jes2WtGGTTc66V/95fef8a946ea9f8b46a951359082e388/c1d7ad1f-7492-4139-9381-6b66ff6731b8.png?w=1600&fm=webp&q=72)

Agent experience metrics are defined as a set in [how to measure agent experience](https://www.agentrail.co/blog/how-to-measure-agent-experience). This page does not redefine that set. It ranks which number to move first when the failure you can see is a weak public score, a lost recommendation, or a failed integration.

Those three failures look related on a slide. They are not the same first fix. A missing OpenAPI file will poison preference tests and task runs for the wrong reason. An unaided competitor pick will not show up as a red readiness grade. A green preference win can still die at auth.

The order follows the failure mode you actually have. The edit after you pick a number lives in [agent experience optimization](https://www.agentrail.co/blog/agent-experience-optimization). The diagnosis of where a task dies lives in [why AI agents fail tasks](https://www.agentrail.co/blog/why-ai-agents-fail-tasks).

> **Key Takeaways**
>
> • There is no universal “always start with readiness.”
>
> • Weak public score: move readiness first, then remeasure before burning preference budget.
>
> • Lost recommendation: move preference first when the public contract is already adequate.
>
> • Failed integration: move task completion and the dead step first when agents already choose you.
>
> • Remeasure the instrument you changed. Do not celebrate a readiness lift as a preference win.

## Rank the failure you have, not a default funnel

A fixed funnel ritual sounds tidy. Scan, then preference-test, then measure completion, every quarter, in that order. Real programs break that ritual as soon as the symptoms disagree.

Microsoft’s AX playbook makes a related point for propensity versus efficacy. Which you start with depends on context. A new product often needs the unnamed-task question first. An established product often needs correct use first ([Designing scenarios](https://microsoft.github.io/agent-experience/playbook/designing-scenarios/)). Platform teams need a third branch. The public contract can be the blocker before either scripted question is worth the runs.

Name the symptom before you name the dashboard. Weak public score means the scan says agents struggle to find, read, or use what you published. Lost recommendation means unaided agents name someone else, or your funnel never sees the choose event. Failed integration means agents that already pointed at you still do not finish.

## Weak public score: move readiness first

If the public contract is incomplete, preference and task tests keep failing for the wrong reason. The agent never reaches a fair choose or finish trial. Move [agent readiness](https://www.agentrail.co/blog/what-is-agent-readiness) first.

The [agent readiness leaderboard](https://www.agentrail.co/agent-readiness) makes the public bar concrete. On the November 3, 2026 public board it holds 160 companies, averages 68 out of 100, and tops out at 92. The grade mix is 20 A, 57 B, 73 C, 8 D, and 2 F. The field read of where find, read, and use break is in [what the Agent Readiness Leaderboard shows is failing](https://www.agentrail.co/blog/state-of-agent-readiness).

A readiness lift still is not proof of success. The October measurement essay is explicit. A declared-surface score is a precondition. It does not predict agent success ([How to measure Agent Experience](https://agentexperience.tech/insights/measure-agent-experience/)). Cloudflare’s 17 April 2026 crawl shows the same gap in another instrument. In its filtered set, `robots.txt` appeared on 78% of sites. Content Signals appeared on 4%. Markdown negotiation appeared on 3.9% ([Introducing the Agent Readiness score](https://blog.cloudflare.com/agent-readiness/)).

First move: close the declared gaps the scan names. Remeasure the grade. Only then spend simulation runs on unaided choice. Otherwise you will pay to rediscover that the agent could not find the current docs.

## Lost recommendation: move preference first

If the public contract is adequate and unaided agents still name a competitor, another scan is the wrong first move. The first number is preference.

That failure often leaves no marketing residue. The developer asks for a job, not a brand. The agent installs and writes code. Your analytics may never record that the question was asked. [Coding agents recommend tools your analytics never records](https://www.agentrail.co/blog/coding-agent-recommendations) owns that blind spot. Preference tests are how you make the choose failure visible on purpose.

[Microsoft’s 6 October 2026 definition](https://developer.microsoft.com/blog/what-is-agent-experience-ax/) calls this propensity. Give the agent a task that does not name your product. Record what it reaches for across repeated runs on a stated agent profile. Hold the model, harness, and prompt still when you change one surface that should affect choice.

First move: script the unnamed tasks, record the names, then change the docs or discovery surfaces that shape selection. Do not jump to deep integration rewrites until agents can find you in the unnamed case. Task completion still matters. It is the wrong first chart when the choose step never happens.

## Failed integration: move task completion and the dead step first

If agents already choose you and still fail to finish, do not open with another preference round. The first numbers are task completion and the dead step.

[Agent task completion rate](https://www.agentrail.co/blog/agent-task-completion-rate) is completed tasks over started tasks, judged by environment state. [Anthropic’s 9 January 2026 note](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents) draws the line with a booking. The agent can say the flight is booked. The outcome is whether a reservation exists.

Production agents are short. In [*Measuring Agents in Production*](https://arxiv.org/abs/2512.04123) (revised 3 February 2026), 68% of deployed agents in the filtered set execute at most 10 steps before a human intervenes. The authors surveyed 306 practitioners. They then focused on 86 production or pilot systems. A short path means the first wall is the whole story. Name that wall with the surface and the action. The diagnostic essay is [why AI agents fail tasks](https://www.agentrail.co/blog/why-ai-agents-fail-tasks). The edit is [agent experience optimization](https://www.agentrail.co/blog/agent-experience-optimization).

First move: fix the named step. Remeasure completion, retries, and path cost on the same profile. A preference win you already have will not repair an auth wall or an unusable error body.

## Remeasure in the same order you fixed

After a change, re-run the instrument you moved. Then check whether the next failure mode is still open.

A readiness lift that leaves unaided naming unchanged was still useful. It cleared the precondition. It was not a preference win. A preference lift that leaves completion flat means agents now choose you and still fail to finish. That is a clean handoff to the dead-step work, not a reason to declare the program done.

Microsoft’s practical rule is to run the same task several times on a recorded profile. Their October essay uses five runs as the working habit. Scope every comparison to the profile you measured. A result belongs to that setup.

[Agent analytics](https://www.agentrail.co/blog/agent-analytics) is where production tasks are read together. We reconstruct journeys agents already attempted across website, docs, APIs, SDKs, and MCPs. Simulations cover the scripted preference and named-product cases. Readiness is included. Public pricing for journeys and simulations is on the [pricing page](https://www.agentrail.co/pricing).

## A short order of operations

Keep the sequence portable. Match the symptom, move one number, remeasure, then open the next branch.

| Symptom | First number | First fix lane |
| --- | --- | --- |
| Weak public score | Readiness grade | Close declared find/read/use/pay gaps, then rescan |
| Lost recommendation | Preference result | Unaided tests on a recorded profile, then change discovery/docs that shape choice |
| Failed integration | Task completion and dead step | Name the surface and action, fix that path, remeasure finish and cost |

If two symptoms show up at once, clear the earlier blocker first. A broken public contract contaminates preference. A lost recommendation contaminates any story about “agents use us.” A failed integration is the priority only when choose is already true.

Selection among the instruments, without the priority rule, is covered in [a recommendation and a finished task are different measurements](https://www.agentrail.co/blog/measure-ai-agent-recommendations). Definitions stay in the measurement guide linked at the top.

## Frequently asked questions

### Should every team start with readiness?

No. Start with readiness when the public score is the open failure. If the contract is already adequate and agents still name a competitor, start with preference. If agents already choose you and fail mid-path, start with task completion and the dead step.

### What if two failure modes show up at once?

Clear the earlier blocker first. Fix declared surfaces before you trust preference results. Fix choose before you treat completion as the whole program. Remeasure after each change.

### Where are agent experience metrics defined?

In [how to measure agent experience](https://www.agentrail.co/blog/how-to-measure-agent-experience). This page only ranks which number to move first under three failure modes.

### When do I hand off to optimization?

When you can name the surface and the action that stopped the task, hand that name to [agent experience optimization](https://www.agentrail.co/blog/agent-experience-optimization). Sequencing picked the number. Optimization changes the step.

## Conclusion

Weak public score, lost recommendation, and failed integration each call a different first fix. Move readiness when the contract is incomplete. Move preference when unaided choice is the hole. Move task completion and the dead step when agents already choose you and still fail.

Keep the metric definitions in [how to measure agent experience](https://www.agentrail.co/blog/how-to-measure-agent-experience). Keep the board patterns in [state of agent readiness](https://www.agentrail.co/blog/state-of-agent-readiness). Keep the choose blind spot in [coding agent recommendations](https://www.agentrail.co/blog/coding-agent-recommendations). Keep the finish diagnosis in [why AI agents fail tasks](https://www.agentrail.co/blog/why-ai-agents-fail-tasks). Then change one thing, remeasure that instrument, and only then open the next branch.