to be clear, this is a closed source orchestrator on top of closed source models. if before you didn't control the models, now you don't even control which ones are used or how much. this is not "AI sovereignty" i've also read the tech report to get an opinion on the technical stuff: fugu (not the ultra version) is basically a classifier that selects which model at each turn is most likely to answer correctly (in other words a router). this leads to -10 points on SWE Bench pro compared to opus, gets some gains on other benchmarks but very slight. argument could be that it reduces cost, but no information about this so it's likely the opposite. they also have an autoresearch benchmark where they compare to frontier models "Model A, B and C" which is really crazy to not be transparent about what models you compare against. let's also say that this probably doesn't support adding new llm out of the box since you need to retrain the classifier about fugu ultra, this is basically and advanced plan mode and orchestrator, this is a model that for a query outputs a plan with multiple "workflows". my understanding of workflows is that they say: "spawn model A subagents to achieve this, then use model B to judge it, then summarize this with model C" which is just a test time scaling compute strategy. i think this is an okish way to do it, but it's limited by the fact that they need to predict everything before the agents start working, which is why they limit this to 5 steps. imo you need to predict what to spawn at t+1 with the information you get at t, not with the info you get at t=0. there are also other issues such as fable 5 score on terminal bench being wrong and them being super vague and unclear about which model is in the LLM pool (they only mention closed source api one) the biggest and most obvious issue is that they are introducing a "test time scaling" method with "best of N" over models, and they literally NEVER REPORT the number of output tokens or cost to achieve a benchmark/task the good comparison here is not with opus, but it's opus with ultracode/workflows enable, not with kimi, but with kimi swarm ect.. very very confusing release

Most software engineers are facing an identity crisis bordering on depression. As CTOs aggressively evangelize tokenmaxxing, a class divide ensues. The lazy. The lazy push code. They don't write it. They don't manually test it. They don't even read it. They're on autopilot. See Jira ticket, prompt for task, submit code. Many of them are barely on their computer the whole day. A comment on the PR asking why they did this? The lazy ask AI. A Slack message? The lazy ask AI. Need to prepare for standup? The lazy ask AI. As long as it sounds enough like them and isn't detected. Some of the lazy are even overemployed, and work multiple jobs. The lazy smart ones get away with this, and even rewarded. After all, software engineering for the lazy is just a dance to convince your colleagues you're smart and hard working. The craftsmen. The craftsmen are tired. Very tired. 15 PRs in queue. Slack blowing up. The entire burden of review falls on the craftsman. The burden of understanding. They try. They work their way through the code, thoughtfully commenting to improve what ships. The response? A lazy: "That's a clever idea! You're absolutely right." with an incorrect change. It's fine, the craftsman says. I can fix them. They write a doc urging his colleagues to be better. The next day? 20,000 line PR to review. Day after day, their workload grows. Bugs seep into production. No one seems to care. Another round of AI is thrown at it. Their animosity to their colleagues rises. Eventually, they give up. It's just not what it used to be. The craft they loved is dead. They eventually wake up, a lazy. This isn't all companies. Many companies are genuinely more productive, adopt the right set of principles and practices around AI development and have highly talented teams that trust each other. It tends to happen in bigger companies that are 10+yrs old with a higher talent variance. But it happens. A lot.

People have been asking why OpenHands V1 goes the opposite direction from Claude Managed Agents. I finally found the time to write it up. The production pain points are real, but changing the hosted agent architecture won't solve them for free:

SkillsBench is a great benchmark, and I'm not just saying that because @OpenHandsDev beats everyone else on it 😀 Skills are now the defacto standard for how we customize agents, so figuring out models/agents ability to use them well is super-important.

🧵 Claude-Opus-4.8 takes you too much tokens - but is this issue general across agents? Do agents know how much they'll spend? Introducing Budget-Aware Agents (BAGEN): We study budget awareness across 4 envs & 5 frontier agents, and find structured failures in most of them. 👇

I’ll be at MLSys 2026 in Seattle this week! We’re presenting our OpenHands Software Agent SDK work — lessons from OpenHands V0, where we ran into the hard parts of scaling software agents to production. If you’re around, come say hi! Happy to chat about agent infrastructure, evaluation, training, verification, or anything software agents. Oral: Thu May 21, 9:30–9:45 AM PDT Poster: Thu May 21, 6–8 PM PDT, Poster Session 3

🚀 The most realistic coding benchmark yet. The latest EvoClaw continuous & long-horizon coding leaderboard is out! This update features Opus 4.7 xhigh (200k/1m), GPT-5.4 xhigh, GLM-5.1, and GLM 5. (GPT-5.5 and Kimi-k2.6 are coming in the next update 👀)

Why are AI agents so expensive? Do more tokens actually lead to better performance? Which models are more token-efficient? Can agents predict their own token costs before execution? These were the questions bugging us, so we wrote a paper to find out. Excited to share "How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks", led by @Longju_Bai, with co-authors at Stanford, MIT, Google DeepMind, Microsoft AI, and All Hands AI. A few findings that surprised us: 🔹 Agentic coding tasks consume ~1000× more tokens than chat or reasoning workloads. And input tokens, not output, become the dominant cost driver, because each round re-feeds the entire trajectory back into the model. 🔹 More tokens ≠ better outcomes. Runs on the same task can vary by up to 30× in token use, and accuracy often peaks at intermediate cost. Beyond that, extra spending tends to reflect redundant exploration and does not bring further performance gain. 🔹 Models differ substantially in token efficiency. On the same successfully solved tasks, Kimi-K2 and Claude Sonnet-4.5 use roughly twice as many tokens as GPT-5.2. The gap becomes even larger when all the models fail. 🔹 Human-rated task difficulty weakly predicts actual cost. "Easy" tasks for humans can be surprisingly expensive for agents, and vice versa. The classic "Moravec's Paradox" is also true for coding agents! 🔹 Agents struggle to predict their own costs. Self-prediction correlations top out around 0.39, and every model we tested systematically underestimates what a task will cost. Result-based pricing still has a long way to go when we cannot even figure out the token cost beforehand. Together, these results suggest that cost prediction is a genuinely challenging task for current agents. We think this opens up real research questions around self-modeling, calibrated cost estimation, and pricing mechanisms that work under residual uncertainty. Huge thanks to my collaborators: @Longju_Bai, Zhemin Huang, @sunjiao123sun_ , @xingyaow_ , @radamihalcea , @erikbryn , @alex_pentland paper: website:

Great work! This pretty much aligns with what I saw in our PR data as well. Agent-authored code is either mostly accepted, with 90–100% survival, or mostly rejected, with 0–10% survival.

Reading this made me appreciate one subtle upside of cloud agents a bit more. The credential setup can be annoying, but it’s also a useful bit of friction. A local agent may accidentally inherit all the random authority sitting on your laptop: env vars, cloud CLIs, ~/.aws, ~/.ssh, forgotten tokens, etc. With a cloud agent, you’re usually much more deliberate about what credentials it gets.