SignalNov 3, 20261 min read6 sources

Agents are being measured on real economic work, and they still fail most of it

In one minute

New benchmarks now measure agents on long, economically valuable tasks. They cover specialist domains, real engineering, everyday personal-assistant work and open math problems. The results are weak: the best model scores 12% across 206 domain tasks, and agents without web access rediscovered at most 7 of 40 solved problems.

A developer survey found that only 18% of teams reach production and 10% reach 90% success, so builders should plan for partial completion.

From the Field Notes:

Evidence / sources

  1. 1newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities
  2. 2gdpx.semianalysis.com
  3. 3xplore.ai/rwe-bench
  4. 4openai.com/index/lifeval
  5. 5arxiv.org/abs/2610.37739
  6. 6github.com/aliyun/ai-agent-handbook/blob/main/2026-agent-survey-report.md
All Signals

Next reads

Read next

Open the graph
  1. 01topicClaudeExplore the claude topic hub.
  2. 02topicClaude CodeExplore the claude-code topic hub.
  3. 03topicCodexExplore the codex topic hub.
  4. 04archiveSignalsOpen the signals and the raw-signal archive.
  5. 05sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract