Kiro's MAESTRO proof harness solves all six 2026 IMO problems in 1h46m

AWS's DevOps blog describes MAESTRO as a harness rather than a model, sitting inside the Kiro CLI developer agent. Its job is to coordinate an LLM through formal proof development. It explores candidate approaches, decomposes a theorem into lemmas, and tracks structured progress, so long proofs don't collapse under context limits or dead ends. AWS reports MAESTRO solved all six 2026 IMO problems in 1h46m and proved more hard theorems than general-purpose coding agents, naming OpenAI's Codex and Anthropic's Claude Code.

The significance is methodological. Results like this used to come from bespoke systems, such as DeepMind's AlphaProof lineage or specialized research harnesses. MAESTRO packages the orchestration in a shipping developer tool, which suggests that scaffolding around frontier models now matters as much as raw model capability for long-horizon reasoning. That echoes Google's Antigravity Orchestrator pairing with Gemini 4 Hydrogen the same week.

The competitive framing deserves scrutiny. Codex and Claude Code are general coding agents and are not tuned for formal mathematics, so beating them on theorem proving is a narrower claim than it sounds. The blog doesn't specify which underlying models MAESTRO used, whether results were verified in Lean or a similar proof assistant end-to-end, or how many attempts were allowed.

Watch for open release of the harness or benchmarks, independent verification of the IMO proofs, and whether the decomposition patterns generalize to software verification. AWS has a long internal history with automated reasoning, and that commercial use case matters more than olympiad bragging rights.