AI Agents: Real Cases of Wins, Failures & Hard Lessons
Updated October 29, 2026 · about 15 minutes Most guides define the term and stop. Method: I read the primary sources where they exist (Gartner, McKinsey, METR, Anthropic) and flagged every figure that comes only from th
Updated October 29, 2026 · about 15 minutes Most guides define the term and stop.
Method: I read the primary sources where they exist (Gartner, McKinsey, METR, Anthropic) and flagged every figure that comes only from the company that benefits from it. In an agent, the model decides the path itself, in a loop, until it thinks the job is done (Building effective agents). How long a task can an agent handle?
Key takeaways
- METR measures the length of software tasks (in human hours) that an agent completes about half the time: | Model | Approx.
- Phase 2, published December 18, 2025, added newer models, a second “CEO” agent, and better tooling.
- Two agents on the same model share the same blind spots.
The METR horizon values above come from trackers summarizing METR, and METR itself warns measurements beyond 16 hours are unreliable.
Read the full article: https://www.best-ai-agents.io/ai-agents-real-cases-of-wins-failures/
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.