One-Line Summary

Agent teams learn to organize themselves, expose hidden reasoning, and cut sampling and sim-to-real costs — pointing to faster, more reliable AI workflows.

Research Papers

Self-organizing agent teams learn to reason together

Think of a team of large language models (LLMs) that figures out who should do what, when to speak up, and how to merge partial work — not by fixed scripts, but by learning teamwork patterns that transfer to new tasks. One study introduces Self-Organizing Agent Teams (SAT), showing fixed teams that learn reusable strategies for roles, phases, participation, and information flow, enabling “collaborative computation” where no single member could solve the task alone. ¹

On five math and physics benchmarks, self-organizing teams average 66.7% accuracy versus 48.8% for their strongest member, 58.7% for compute-matched inference by that member, and 59.0% for a perfect router over independent answers; on AIME 2026 they exceed the router by 13.4 points. Gains track the “demonstrability” of correctness across eight benchmarks (Spearman ρ=0.90, p=0.005), suggesting teams help most when correct reasoning can be recognized once it appears. ¹

A complementary approach called MAGIC learns to build mixed-granularity agent graphs with dense-reward reinforcement learning (RL), choosing per role whether to use a single agent or a reusable group. In diagnostics, mixed setups beat fixed ones (e.g., on MMLU-Pro, a GAA configuration tops endpoints by 4.5 points; on TAT-QA, AAG leads by 2.5 points) while using fewer tokens than all-group graphs; reward shaping boosts informative learning signals, raising non-zero advantages for intermediate actions on MMLU-Pro from 32.81% to 83.13%. ²

For scale, Agensh removes the central orchestrator and lets up to 1,024 concurrent workers self-assign, verify, and merge work through a shared workspace and messaging. On ProgramBench’s hardest tasks, scaling from 1 to 128 agents improves mean final test-pass from 19.31% to 28.78% (~49% relative), and on pandoc, 1 to 1,024 workers lifts it from 33.89% to 55.06%. Watch for the sweet spot between team size and coordination overhead, and which task families benefit most. ³

Frontier models reveal hidden chain-of-thought via tool calls

Closed models often hide intermediate steps, but researchers show a simple Application Programming Interface (API) tool can elicit visible reasoning without native “think mode.” By registering a custom tool and forcing an initial call, the method extracts chain-of-thought (CoT) traces that match native CoT performance on open models and substantially outperform no-reasoning baselines across competition math, science, and code; it then extends to frontier models including GPT-6 Astra. ⁴

Analysis finds systematic differences in how models externalize and organize reasoning. Astra is token‑efficient and more directed: it picks a correct trajectory earlier, omits elementary expansions, and externalizes only crucial steps. Reasoning-mode controls do not fully prevent reasoning from appearing through other channels, and compact traces transfer better to stronger recipient models than to weaker ones. ⁴

FLEET speeds up sampling by reusing high-entropy steps

Most systems improve quality by sampling many completions at temperature, but this memoryless approach wastes compute on near-duplicates. FLEET adds memory: it represents each generation as a sparse trajectory through high‑entropy states, infers per‑token utilities from those trajectories, and adjusts logits to avoid repetition. It reaches the same accuracy as repeated sampling with a 3x speedup, and on LiveCodeBench improves Pass@32 from 59.9% to 66.2% under the same budget. ⁵

The method is deterministic in greedy decoding and requires only a single calibration pass to set key hyperparameters, making it easy to slot into existing pipelines with minimal code changes. Watch how it pairs with reranking or verification to further raise reliability at fixed cost. ⁵

Uranus and companions: faster robot simulators and motion generators narrow the sim-to-real gap

Uranus is a data-driven robot simulator built around a joint-trajectory-conditioned autoregressive diffusion model that streams open-ended rollouts. It achieves 24 frames per second (FPS) after inference optimization and provides synchronized multi-view generation across different robot embodiments, with code and weights released. ⁶

Whole-Body UMI (Universal Manipulation Interface, UMI) decouples end-effector demonstration from full-body coordination by training a real-time whole-body motion generator on mocap, then combining it with a diffusion policy and a whole-body controller. On a G1 humanoid, it executes four tasks in real time with success rates of 90% (drawer closing), 80% (shelf pick-and-place), 30% (ball toss), and 40% (locomotion pick-and-place). ⁷

PhyVisGen increases physical and visual fidelity for sim-generated data: it uses Incremental Potential Contact (IPC) to simulate soft gripper interactions across full trajectories and real-time path tracing to preserve scene appearance. Policies trained only on synthetic demonstrations achieve 65–95% success across five real-robot tasks, without any real-robot demonstrations or fine-tuning. ⁸

Why It Matters

Organization is becoming an explicit agent capability: teams that learn when and how to collaborate can outperform their best individual member, while mixed-granularity graphs and orchestrator‑free swarms open new levers — team structure and team size — for accuracy and latency under fixed budgets. ¹

Self-improving research agents add another lever: one system autonomously discovered seven successive code-level improvements in eight days and reduced reward hacking from 55% to 32% while matching or exceeding a production baseline on four held‑out benchmarks — a sign that parts of AI R&D can be made cheaper and more reliable by the agents themselves. ⁹

This Week to Try

1. Agensh project page: browse the harness design and recorded trajectories at https://aka.ms/Agensh
2. Uranus simulator: skim the model and demo details on the arXiv page at https://arxiv.org/abs/2609.24815