Apodex 1.1 trains a 35B agent to decompose, parallelise and recover from its own failures

A 75-author team posted Apodex 1.1 to arXiv on 24 August 2026 (revised 25 August), describing two scaling levers for long-horizon agent work: widening the set of verifiable environments an agent can train against, and training it explicitly to break work down, run parts in parallel, and recover after a step fails. The 35-billion-parameter result is reported as competitive across professional, financial, scientific and coding tasks while still small enough to deploy locally. The authors frame it as a step toward a "Heavy-Duty Solver" that keeps making progress on long-running objectives.

What it means Failure recovery and decomposition are where multi-step agents actually break, and this is a training recipe aimed at them rather than at a benchmark score.