Generative AI Fundamentals Map
Map/Loops/reinforcing

Trust and delegation

reinforcing loop · 4 steps · Trace on the map →

The cycle

Agents prove reliableon real tasksMore work delegatedMore audit evidenceReliability improvesR
  1. 01Agents prove reliable on real tasksReal-world task evals
  2. 02More work delegatedAgent-native organisations
  3. 03More audit evidenceAudit and attribution
  4. 04Reliability improvesReliability on repeat

What drives it

Evidence of reliability earns more delegation, and more delegation produces more evidence.

What breaks it

One visible failure can reset trust. METR found developers felt faster with AI while measuring 19% slower, so perceived and real reliability can diverge.

What to watch

pass^k on real workflows. Share of tasks run without human review.

Also feeds it

Sources

  1. 01τ-bench: tool-agent-user interaction, and pass^k
  2. 02Import AI 429 on GDPval
  3. 03METR: AI coding can make developers slower (The Decoder)