← The Wire

Public story · 2026-06-30 · high

Confident AI's guide grades agents on trajectory, not answers

Agents can land on the right answer through an unsafe path, and scoring only the output misses it.

Why now: This lands as one specific fix for a known blind spot: scoring answers without ever checking how an agent got there.

Confidence
high
Sources
1
Redaction
passed

Story

Agent-as-judge evaluation checks an AI system's intermediate steps and tool calls, not just its final output, per a guide from Confident AI.

That catches the failure output-only scoring misses: an agent that reaches the right answer through a wrong or unsafe path. In production, that's the run that blindsides a team. It looks clean. It wasn't.

The method runs a cheap agent-as-judge on 100% of traffic: every trace, every tool call, every intermediate step, scored automatically. Teams then pull a small human-reviewed sample, 1-2% of traffic, and check it against the judge's calls. If the two agree, the automated grading holds up.

The 1-2% human sample is the part most teams will cut first. When teams skip it, the judge's blind spots go unchecked: agents start passing on the exact steps a human reviewer would have flagged. Watch for teams that report a judge's pass rate but never publish the human-agreement number. That's the tell they skipped the check.

This lands as one specific fix for a known blind spot: scoring answers without ever checking how an agent got there.

Related stories

Each link below shares sources, entities, or timing with this story.

  1. Shared topic / Tension

    OpenAI Responses API Gets Full Computer Environments.

    Overlapping topics (agent, call, final, intermediate); pushes against this story (vs).

  2. Shared entity: Output / Earlier coverage / Tension

    84% of Developers Use AI Tools. 3% Strongly Trust the Output. Read That Again.

    Both cover Output; earlier Output coverage from 2026-04-02; pushes against this story (but).

  3. Shared topic / Tension

    MCP tool poisoning fires silently on every call

    Overlapping topics (agent, call, catch); pushes against this story (but).

  4. Audit your eval harness against the five-rule checklist.

    Overlapping topics (agent, answer, evaluator); pushes against this story (against).

  5. The Red Queen Gödel Machine co-evolves agents with their own evaluators.

    Overlapping topics (agent, answer, evaluator); pushes against this story (against).

  6. Shared topic / Downstream implication

    Evaluator bias is contagious in multi-agent systems.

    Overlapping topics (agent, evaluator, judge); traces where this leads (downstream).

  7. Shared topic / Tension

    Gate agent CI on pass^k consistency, not a single green run.

    Overlapping topics (actually, agent, human); pushes against this story (against).

  8. Computer-Use Agents Cost 45x More Than APIs. The Numbers Are Brutal.

    Overlapping topics (agent, call, human); pushes against this story (against).

Source trail

Entities

Claim evidence

  1. Confident AI's guide grades agents on trajectory, not answers

Provenance

Canonical issue
2026-06-30
AI generated
yes
Story unit
2026-06-30-score-the-trajectory-not-just-the-final-answer
Labels
source-backed, canonical briefing excerpt