Claim
Long-running, self-correcting AI agents operating without human checkpoints compress months of engineering work into hours, at a cost measured in hundreds of dollars of token spend rather than weeks of salary.
Mechanism
Self-correction eliminates the human-checkpoint bottleneck. A human-supervised sprint requires review at each step; the agent checks its own output, catches errors internally, and continues running. Operating continuously for 14 hours without intervention, the agent accumulates output that would require a team sprint to match. The cost ceiling is the token price per inference call, not the engineer's hourly rate. When token costs run at roughly $18 per hour of autonomous operation, a two-week engineering sprint becomes a $251 overnight run.
Conditions
Holds when: the task can be specified upfront clearly enough for the agent to work without mid-run human input; the output is verifiable at the end against defined criteria.
Fails when: the task requires tacit judgment calls mid-run that cannot be anticipated in the initial brief; when intermediate output determines the next step in ways the agent cannot evaluate alone; or when the cost of reviewing large-volume agent output exceeds the generation cost.
Evidence
Mollick cites Epoch AI research on Opus 4.7 in the June 30, 2026 essay:
"Opus 4.7, working on its own for 14 hours, was able to build a software package that would take 2-17 weeks of human engineering work (it cost $251 in tokens)"
Fable accomplished "complex software projects that would have taken a team well over a week to do" in 9 autonomous hours.
Signals
- Agent build runs have shifted from minutes to hours without requiring human direction between steps.
- Morning code review covers pull requests filed by agents running overnight.
- Token spend per project is measurable and predictable; labor hours per project are not.
Counter-evidence
The 2-17 week range is wide. The actual compression depends on task structure and model choice. Self-correction is not perfect: long-horizon runs can produce large volumes of plausible-but-wrong output that requires costly expert review. If the review and correction time exceeds what building from scratch would have cost, the economics invert. The $251 figure is also model- and task-specific; complex architecture decisions may require more inference steps and higher costs.
Cross-references
- Frontier models capable of multi-hour autonomous runs require commissioning, not steering: the Fable 5 case study from June 9 showing the same pattern at 9.5 hours, which predates and contextualizes this finding.
- When AI agents prompt other AI agents, throughput scales with compute rather than headcount: the infrastructure that makes this possible, where one agent manages subordinate agents across a multi-hour run.