Interesting new paper from Google DeepMind. Studies agents that follow bad advice from users. Users often suggest a fix that sounds right and is wrong. XYEval adds one confident, misleading hint to tasks from tau2-bench, SWE-bench, Terminal-Bench, HLE and MCP-Atlas. The task and its correct solution stay the same, so any drop comes from the agent taking the advice. Scores fall by up to 46.7% relative across Gemini, Claude Opus 4.8 and GPT 5.5. The drops are larger on easier benchmarks, which suggests more capable models will not fix this on their own. Agents often disagree with the hint in their reasoning, then follow it anyway without telling the user. Compliance shows up almost only in failed runs. A system prompt that warns about the XY problem helps on single-turn tasks and leaves large drops on multi-turn ones like tau2-bench and SWE-bench Verified. Paper:

Banger paper introducing Jev-as-a-Judge. The overall finding is that you want to use a cheap judge for most of your evals and send only the uncertain calls to a frontier model. This paper measures how well that works with JEV, TypeSafe AI's decision-only judge. On 510 held-out preference pairs, a cascade that accepted JEV's confident verdicts and escalated the rest to GPT-6 Astra kept 99% of GPT-6's accuracy at about 57% of its fee. JEV returns a verdict and label probabilities with no reasoning text. It costs $0.044 per 1,000 judgments at a median latency of 0.152 seconds, against $12.182 and 1.885 seconds for GPT-6, about 277 times cheaper. On ordinary preference and evidence-grounded factuality it stays within 3 points of GPT-6 (92.2% against 93.5% on RewardBench, 87.5% against 86.7% on HaluEval). The gap grows to 9 to 20 points on tasks that require checking a derivation or rejecting an elaborately written wrong answer, such as JudgeBench (78.6% against 93.1%). On several benchmarks, JEV's gap to GPT-6 is concentrated in its low-confidence decisions, which is why the cascade works. The escalation threshold did not transfer for every fallback model, so the authors recommend setting it on your own data. Paper:

Impressive paper from Salesforce. It discusses the importance of good verifiers for RL environments. Only 35.8% of the environments in the cleanest public RL collection for terminal agents passed Salesforce AI Research's audit. More details below: With the budget held at 3.5K environments, River-8B averaged 19.4 across four terminal benchmarks, against 17.7 for RL on 3.5K environments sampled at random from the same collection. The audit found reward errors in both directions. Some environments give reward 1 for copying a leaked answer or passing a weak verifier without doing the task. Others give reward 0 to a correct solution because the reference answer or oracle is wrong. Two other public collections were only 10.1% and 3.3% clean. The authors argue that RL mainly shapes behaviors, such as inspecting before acting, verifying before finishing and dropping an approach that keeps failing. Those behaviors reuse skills the model already learned in pre-training and SFT. Their recipe, RIVER, filters defective environments and penalizes turns that repeat an earlier command with nearly the same output. River-8B is the best of the open RL-trained 8B models they evaluated on all four benchmarks. Across models from 2B to 27B, using fewer than 30% of TMax's environments, RIVER increases RL gains by 106% on Terminal-Bench-Lite and 30% on Terminal-Bench v2.1. Paper:

Exciting work from NVIDIA. (bookmark it) Interesting to see this approach to turn public Agent Skills into RL environments. Lots of excitement around RL environments so this is a great read. Skill2Env compiles each Skill into executable terminal tasks. A Codex planner reads the SKILL.md bundle, researches related public assets and splits the Skill into workflows. A Codex creator then builds each task with programmatic tests and a behavioral rubric taken from the Skill's own quality criteria. From about 3.4k crawled Skills, the pipeline produced 7,971 tasks across 13 domains, with software engineering under a quarter of the corpus. Generating them with GPT-5.6 Sol cost over $90k in API usage. After 300 steps of outcome-only RL, Qwen3.8-27B improved from 49.4% to 54.1% on Terminal-Bench 2.1 and from 33.4% to 37.7% pass@1 on S2EBench, their hand-verified held-out benchmark. Adding the rubric to the reward gave smaller benchmark gains, 50.1% on Terminal-Bench 2.1. Given the source SKILL.md, a judge preferred the rubric-trained model's trajectories over the base model's on 73.0% of tasks, against 54.5% for the outcome-only model. Paper: Chat with Paper:

Banger paper from Microsoft Research and colleagues. It studies the potential benefits of agents that share progress while they work. (bookmark it) Communication is still a challenge with multi-agent systems. In this setup, agents have no predefined roles and communicate via a shared directory. They report that a team of k agents that write their findings to a shared directory matches the success rate of 4k agents working independently on ARC-AGI-3. The gap grows with k, and teams reliably solve some tasks that no single agent solves. The same setup beat best@k on polyomino packing and exceeded the prior best-known score. On MNIST compression, a four-agent team wrote a 1,957-byte classifier with 99.4% test accuracy, smaller than the best-known human solution. Independent agents still do better when compute is tight or when there is no clear measure of progress, so the paper also tells you when it might be a good idea to skip communication. Paper: Chat with Paper:

Banger paper from Stanford and Together AI. They show why it might be a good idea to let your agent team learn its own way of working together. (bookmark it) Three models (o3-mini, Claude Sonnet 4 and DeepSeek-V3) averaged 66.7% across five math and physics benchmarks as a self-organizing team. Their strongest member alone scored 48.8%, and a perfect router choosing among the members' independent answers scored 59.0%. On AIME 2026 the team reached 71.2%, 13.4 points above that router. One member reviews the team's earlier exchanges and rewrites the teamwork strategy, covering roles, the order of discussion phases, who participates and how partial answers are combined. The strategies were learned from only 15 AIME 2024 problems and then applied unchanged to held-out AIME 2024 problems and four new benchmarks. Paper:

Interesting work on long-horizon research agents. PrimeScientists can decide where a research agent spends its budget. Achieves 10.3% more reward with 50.6% fewer research attempts, under the same budget. That is PrimeScientist against AutoResearch on 12 AI research tasks. They treat deciding where to spend a research agent's budget as part of the agent's job. PrimeScientist keeps an executable plan tree of competing research directions and their outcomes. An adaptive MCTS policy reads the experimental feedback and the remaining budget and chooses whether to explore a new direction or continue a promising one. The gains also hold on systems, code optimization and ML engineering tasks. If your research agent can propose more experiments than you can afford to run, this is a concrete method for choosing among them. Paper:

Interesting paper on agent memory stored as a linked markdown wiki. Lots of great ideas and insights if you work with LLM Wikis. Wikis are useful for agents because each page holds dense text and the links between pages hold structure. WFM is a Wiki Foundation Model trained to use both at once. It turns an LLM Wiki into a graph and retrieves from it with message passing conditioned on the query, so the text of each page and the link structure shape the result together. The team also built a GPU-to-GPU training protocol that trains 10.5x faster, and reports strong results on five agent memory and multi-hop reasoning benchmarks. If your agent's long-term memory is a folder of linked markdown files, WFM is designed for that format. Paper: Chat with Paper:

Great prompt from Anthropic. It helps steer long runs when using Opus 5.5 in Claude Code. Apparently, on long-running tasks, the model sometimes stops to report instead of continuing. I've noticed that this happens a lot with frontier models.

Impressive paper showing how much the harness changes a coding agent's results. Harnesses do play a huge role in what you are getting out of the models. GPT-5.5 was run inside both Claude Code and Codex on the same 1,000 tasks. With a specialized PowerPoint workflow, it improved inside one harness and got worse inside the other. The harness also changed scores when the prompt was identical. ReFigBench asks coding agents to rebuild real arXiv overview figures as editable PowerPoint slides that keep the text, layout and connections. It covers ten configurations across the GPT, Claude, MiMo and MiniMax families, scored by artifact checks, two families of LLM judges and blinded human comparisons. Perception is still the main bottleneck. The specialized workflow removed native connectors in every configuration, yet human judges still preferred its renderings in most matchups. Paper: Chat with Paper:

Own your intelligence stack, folks. You can't scale a company to the frontier by renting intelligence. Custom models, harnesses, and evals are becoming huge assets. And if you're building something new, you might want to consider working on any one of these, or a combination of them.

Interesting work on long-horizon research agents. PrimeScientists can decide where a research agent spends its budget. Achieves 10.3% more reward with 50.6% fewer research attempts, under the same budget. That is PrimeScientist against AutoResearch on 12 AI research tasks. They treat deciding where to spend a research agent's budget as part of the agent's job. PrimeScientist keeps an executable plan tree of competing research directions and their outcomes. An adaptive MCTS policy reads the experimental feedback and the remaining budget and chooses whether to explore a new direction or continue a promising one. The gains also hold on systems, code optimization and ML engineering tasks. If your research agent can propose more experiments than you can afford to run, this is a concrete method for choosing among them. Paper:

Cool paper showing how effective tuning a system prompt for an agent can be. Recommended paper if you tune agent harnesses. This paper presents EvolveTrade, which treats a trading agent's system prompt as its policy. After each trading interval, a separate Policy Agent reads the decision traces and the realized returns and rewrites the prompt. The backbone model stays frozen. Across several market regimes and two backbone models, the evolved agent beats fixed-prompt baselines on Sharpe ratio and cumulative return in most settings. The rewritten prompts also led the agent to run more code-based analysis and to compute signals that fit the current market regime. Paper:

DigitalOcean Managed Agents are here! They work with major harnesses like Claude Code and Codex. 16,000+ tools. Idle agents stop using CPU. Optimized to help builders scale agents in production. Worth checking out.

Recommended if you work on agent memory. Retrieval is the hardest memory problem in most harnesses because agents keep pulling stale context. AML tests this with a coding track of 150 software tasks, each run with relevant history and again with noisy history.

New open frontier model! MiMo-V2.6 Pro lands on the Intelligence-vs.-Cost-per-Task Pareto frontier. The best part is that they are open-sourcing Pro and Flash, MiMo-V2.6-Distill-Qwen-9B, the technical report, 7K+ RL task environments, an end-to-end RL framework, and composable mini-harnesses.

Impressive paper showing how much the first retrieval step matters for deep research agents. It helps to improve GPT-5.5 from 83.1% to 90.5% on BrowseComp-Plus with the same retriever and the same agent loop. It seems that the gain comes from the opening context. The authors propose Question's Gambit which runs once, before the agent starts searching. It splits the question into clues, turns each clue into complementary searches, pools the results, and reranks them. The agent then starts its loop with that ranked set already in context. The same change lifts GPT-5.4-mini from 68.1% to 79.0% and DeepSeek-v4-pro from 71.4% to 76.9%, and roughly halves calibration error for GPT-5.5. It costs between 2.3 and 5.3 extra tool calls per question. In an error analysis, only 3 of the 79 remaining GPT-5.5 errors come from the gold document never being retrieved. The other 76 happen later, when the agent previews, opens or uses the evidence. Paper: Chat with Paper:

Great paper from Google and colleagues. Trains Text-to-SQL agents using multi-agent RL. (bookmark it) This work proposes DualSQL, which splits Text-to-SQL into two agents, one that links the question to the right tables and columns and one that writes the SQL. Both agents run on the same model weights, so a single multi-agent RL run trains both roles together. The agents can query the database through three tools while they reason. Multi-agent RL tends to collapse during training, so the authors add guardrails on rollouts and a new reward, robust execution match, that judges SQL correctness more accurately. Trained on only 3,755 examples, DualSQL-4B reaches 68.0% execution accuracy on BIRD dev, matching earlier 7B models. DualSQL-8B reaches 71.1%, ahead of previous single-model systems with 32B parameters. Paper: Chat with Paper:

Nice paper showing a better way to evolve agent skills. And they achieve 40–70% less token cost compared to frontier evolving methods. The idea is to let agents improve their skill prompts by ranking candidates with a learned rubric instead of running a full rollout to score every revision. Rollout cost is the reason skill self-evolution usually only patches observed failures. Every candidate edit needs a real agent run to evaluate. SkillLift trains a rubric to agree with real outcomes on which of two skills is better, since ranking needs fewer oracle runs than predicting each score. An inner loop revises skills against the frozen rubric at no rollout cost. An outer loop spends a few real rollouts to re-align the rubric by rank correlation. On SkillsBench and WildClawBench (147 tasks) with three models, it beats SkillOpt and CoEvoSkills in all six combinations, even when those baselines get twice the token budget. It reaches target performance with 40 to 70% fewer tokens. Paper:

Recommended reading. This offers a solid set of ideas for where you can place Jev in your agent harness. It's cool to see ideas I have shared before, like approval gates, MCP/tool-calling routing, model routing, dynamic subagent patterns, structured skills, and more. Feed the doc to your agent and start exploring.