Are the agents accomplishing tasks — not just running fast and cheap?

Latency and cost say how fast and expensive a model is. This asks the question those miss: did the agent actually accomplish the goal? Success, human-intervention, and rework rates — per agent and per workflow.

Bottom line

Fleet success 89% across 1212 tasks — but only 63% of agents emit a measured signal (2 inferred, 1 dark). 8% of tasks needed a human. Biggest lift: support-copilot (~56 failed tasks).

Do nextCompare speed & cost →Fix support-copilot first →

Outcome signal — known vs guessed

Outcome signals need run-outcome instrumentation. 5/8 agents emit a measured signal; 2 are inferred from retries / abandons / errors; 1 are dark (shadow). Values are projected where signal ≠ measured — treat inferred rates as directional, not exact.

89%

fleet success rate

1212 tasks, signal-bearing

needed a human

intervention / handoff

13%

rework rate

retried / redone

quality score

63% measured coverage

Fix this first — by failed-task volume

tasks × (1 − success) · biggest absolute lift

Where the most goals are actually missed — a high-volume agent at 82% can fail more tasks than a low-volume one at 48%. Fixing the top row moves the most outcomes.

support-copilot measured

82%~56 failed marketing-gpt inferred

48%~33 failed data-pipeline-agent measured

97%~16 failed Codex CLI inferred

66%~13 failed Claude Code measured

93%~10 failed Cursor measured

90%~10 failed Claude Desktop measured

95%~1 failed

By workflow

WorkflowTasksSuccessIntervene

data-syncmeasured54097%2%

support-triagemeasured31082%12%

code-reviewmeasured26092%7%

content-geninferred6448%34%

infra-opsinferred3866%21%

By agent

AgentTasksSuccessInterveneQuality

support-copilotmeasured31082%12%79 marketing-gptinferred6448%34%—data-pipeline-agentmeasured54097%2%91 Codex CLIinferred3866%21%—Claude Codemeasured14293%6%88 Cursormeasured9690%8%84 Claude Desktopmeasured2295%5%90 Zed Agentnone————

This is task success, not speed — latency & cost live in Performance →. Derived: effectivenessCoverage · fleetEffectiveness · lowestEffectiveness — from app/data/c16-fleet.ts (proposed run-outcome signal; measured / inferred / none provenance per agent).