New Technique Boosts AI Coding Agents by 12% — Just by Thinking Smarter, Not Harder

What if AI coding assistants could get dramatically better without any retraining — just by spending more time thinking?

A new paper from Anthropic tackles a fundamental problem: when AI agents work on complex coding tasks involving hundreds of steps (reading files, editing code, running tests, fixing errors), traditional methods of "try multiple times and pick the best" simply break down. The trajectories are too long and complex to compare directly.

The researchers propose two elegant solutions:

**Recursive Tournament Voting (RTV)** — Run multiple attempts in parallel, compress each into a structured summary, then pit them against each other in a tournament-style bracket until the best solution wins.

**Parallel-Distill-Refine (PDR)** — After each attempt, distill the key lessons (what worked, what failed) and feed them into the next attempt. The AI learns from its own mistakes without starting from scratch.

The results speak for themselves:

• On real-world bug-fixing benchmarks: accuracy jumped from 70.9% to 77.6%
• On the hardest coding challenges: a massive leap from 46.9% to 59.1% (+12.2 points)

The breakthrough insight is treating this as a representation problem. Long agent trajectories need to be compressed into useful summaries that preserve critical information while cutting noise — much like how a senior engineer would review a junior's debugging session and extract the key takeaways.

This matters because it means existing AI models can be significantly improved at deployment time, without waiting for the next generation of models. Just let them think longer and smarter.