← The Wire

Skills2026-03-16 · source-backed

Outcome-First Agent Grading.

Confidence
source-backed
Sources
1
Redaction
passed

Story

Grade final environment state (does the database record exist?), not tool-call sequences. Agents find valid alternate paths — transcript grading creates false failures. Build separate transcript vs. outcome eval layers to expose hidden failure modes. Anthropic Engineering

Related stories

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

  1. Shared entity: Anthropic Engineering / Same source / Tension

    Pass@k vs Pass^k: Know Which Metric You Need.

    Both cover Anthropic Engineering; cite the same source (Anthropic Engineering); pushes against this story (vs).

  2. Shared entity: Anthropic Engineering / Same source domain / What happened next / Tension

    Opus Orchestrator + Sonnet Workers Pattern — 90.2% Quality Gain.

    Both cover Anthropic Engineering; reported by the same outlet (anthropic.com); picks up the Anthropic Engineering thread on 2026-03-19.

  3. Shared entity: Anthropic Engineering / Same source domain / Earlier coverage / Tension

    Sub-Agent Compression Pattern for Token Savings

    Both cover Anthropic Engineering; reported by the same outlet (anthropic.com); earlier Anthropic Engineering coverage from 2026-02-27.

  4. Shared entity: Build / Same source domain / Earlier coverage / Tension

    Defend Against AI-Orchestrated Espionage

    Both cover Build; reported by the same outlet (anthropic.com); earlier Build coverage from 2026-02-25.

  5. Shared entity: Build / Shared topic / What happened next

    Microsoft made agent identity a platform primitive with "Autopilots" and Scout.

    Both cover Build; overlapping topics (agent, environment); picks up the Build thread on 2026-06-08.

  6. Microsoft Execution Containers enter preview.

    Both cover Build; overlapping topics (agent, environment); picks up the Build thread on 2026-06-07.

  7. Shared entity: Anthropic Engineering / Same source domain / What happened next

    Compaction: distill your agent's context before it hits 100% and cut tokens 84%

    Both cover Anthropic Engineering; reported by the same outlet (anthropic.com); picks up the Anthropic Engineering thread on 2026-06-04.

  8. Anthropic Multi-Agent Tiering: Opus Orchestrator + Sonnet Workers = 90.2% Accuracy Improvement.

    Both cover Anthropic Engineering; reported by the same outlet (anthropic.com); picks up the Anthropic Engineering thread on 2026-03-19.

Source trail

Entities

Provenance

AI generated
no
Story unit
2026-03-16-outcome-first-agent-grading
Labels
source-backed, canonical briefing excerpt