AI agent observability · for engineers debugging a run that went wrong

Was it the agent, or who got the work?

Agent traces show what an agent did. They don’t show why it got the job, or who else could have. Paste a decision record below and get a plain answer: which layer failed, and what to change.

Read a record How matching works

Runs in your browser; nothing is sent
Reads the console’s and the API’s decision records
Free, no sign-up

A decision record in the Fivexer console: four candidates with their score, priority and reasons, one chosen and one excluded by a site veto. — The real console. One decision: who was chosen, who was eligible, who was ruled out and why.

The real console. One decision: who was chosen, who was eligible, who was ruled out and why.

Read a decision record

Pick one of three incidents, or paste your own record. The reader names the layer that failed and the setting to look at.

The record

{
  "id": "dt_8f2",
  "taskId": "task_refund_311",
  "workerId": "agent-refunds",
  "mode": "best-match",
  "candidates": [
    {
      "workerId": "agent-refunds",
      "eligible": true,
      "chosen": true,
      "score": 2,
      "effectivePriority": 8,
      "reasons": [
        {
          "kind": "tagWeight",
          "tag": "refund:small",
          "weight": 1
        },
        {
          "kind": "tagWeight",
          "tag": "lang:en",
          "weight": 1
        }
      ]
    },
    {
      "workerId": "payments-2",
      "eligible": true,
      "chosen": false,
      "score": 1.6,
      "effectivePriority": 7.6,
      "reasons": [
        {
          "kind": "tagWeight",
          "tag": "refund:small",
          "weight": 0.8
        },
        {
          "kind": "tagWeight",
          "tag": "lang:en",
          "weight": 0.8
        }
      ]
    },
    {
      "workerId": "payments-5",
      "eligible": false,
      "chosen": false,
      "score": 0,
      "effectivePriority": 0,
      "reasons": [
        {
          "kind": "veto",
          "tag": "refund:small",
          "pattern": "refund:*",
          "source": "manual"
        }
      ]
    }
  ]
}

An example record. Worker names are made up. Choose “Paste your own” to edit.

Which layer failedRouting rules

agent-refunds won, as configured.

It reached 8.0 against payments-2 at 7.6. If someone else should have had this job, the rules sent it here: change the rules, not the worker.

Everyone recordeddecision record · mode best-match, highest priority wins

agent-refundspriority 8.0+1 for refund:small · +1 for lang:enChosen

payments-2priority 7.6+0.8 for refund:small · +0.8 for lang:enEligible

payments-5not rankednever offered refund:small (a veto you set)Ruled out

What to change

Lower the weight on the tag that decided it, or veto that tag for this worker.
Send the same job again in the playground and check the new winner.

Where each answer lives

Keep this beside you when a run goes wrong. Join the two records on the task ID; the rest is yours to log.

The question	Where the answer is	Recorded by Fivexer
Who got the job, and why them?	Decision record: chosen worker, score, priority, reasons	Yes
Who was ruled out, and by which rule?	Decision record: vetoes, jobs turned down, cooldowns	Yes, up to a cap
Did the worker say it succeeded?	Task result: success, error, output	Yes
Which model calls, tools and tokens did the agent use?	Your agent’s own tracing	No, you log it
Did the customer actually get the result?	Your own acceptance check	No, you check it

A decision record holds the people evaluated for that job, up to a configured number. Someone who was never a candidate does not appear in it.

The checklist, in order

1 · Who got it

Open the decision record

Chosen, eligible, ruled out. If the wrong worker won, the rules decided it: weights, vetoes, skills.

2 · What it did

Open the agent’s trace

Same task ID. If the right worker got it and the result is wrong, the cause is inside the run.

3 · Was it done

Check the outcome

A finished task can report success: false. Route that to a person so it is never counted as done.