The agent trust problem is everywhere today, and it's showing up in three different forms at once: Anthropic is building new plumbing for agent coordination, a prominent AI educator admits he still won't ship without manual review, and OpenAI is quietly sweetening the deal on its most capable model. The through-line from yesterday's digest holds: agents are running more, humans are controlling less, and the industry is working out the accountability gap in real time.
01
Anthropic's answer to "I'm the bottleneck" is a new kind of agent workflow
At AI Engineer New York, Anthropic's CJ Avilla and Harrison Stall unveiled dynamic workflows, or "flows" — a new multiagent coordination feature built on top of Claude Managed Agents. The framing was unexpectedly honest: Avilla described the problem they're solving as developers drowning in agent-generated pull requests, stuck approving work instead of doing it. The architectural inspiration they cited was Unix and Linux, where layers of trusted maintainers replaced single points of control. Flows started in Claude Code and are now expanding outward. ---
Why it matters: Garry Tan's workflow from yesterday's digest — approve the plan, let it run — is one person managing a handful of agents. Flows are Anthropic's bet that the next step is dozens of agents managing each other. If it works, your senior developers stop reviewing code and start reviewing agent hierarchies. That's a different job, and not everyone will want it.
OpenAI drops long-context fees for GPT-6.1 Sol subscribers
Thibault Sottiaux, who works at OpenAI, confirmed that long-context usage for GPT-6.1 Sol is no longer hitting subscribers with extra fees. No blog post, no announcement — just a note that generated 2,500 likes and 157 replies, which tells you how many people were quietly absorbing those charges. ---
Why it matters: Long-context fees were the hidden tax on any serious document workflow. Legal teams running contracts through Sol, analysts feeding in earnings reports, developers giving the model full codebases — all of them were getting dinged. Removing that friction for subscribers makes the professional use case significantly cleaner.
"I'm not at a point where I trust the agent, especially with stuff people are using"
Peter Yang shared a quote from Micky, who has taught over 100,000 people to build with AI. The approach Micky describes: let the agent do all the work, collect as much evidence as possible about what it did and why, and only then do a final manual review. He's not skipping the review step — he's redesigning it so it comes at the end with maximum information. ---
Why it matters: This is what "human in the loop" actually looks like in the hands of someone who builds AI products for a living, not a compliance deck. The review doesn't go away. It moves to the end and gets more evidence to work with. If you're deploying agents in your product and calling the approval step optional, you're skipping the part that the most experienced practitioners still won't skip.
Jack Clark's Import AI covers a new benchmark called Humanity's Sixth Sense, built by Elorian and Scale AI, that tests whether AI models can reason about what a visual scene *implies* rather than what it shows. Can the model look at a bookcase and judge whether more books fit? Can it tell whether a chain is bearing a load? Humans score 93.1 on the benchmark. The best AI system, GPT-6 Astra, scores 53.6. Chinese open-weight models fare worse, with Kimi K3 at 25.5. ---
Why it matters: The gap matters most for any product that uses vision AI to make physical-world judgments: warehouse robots, construction monitoring, medical imaging, autonomous vehicles. A model that misses the implied physics of a scene roughly half the time is not a model you deploy unsupervised in those contexts.
Microsoft is building AI students to train better AI tutors
Microsoft Research published details on StudentSim, a system that simulates individual learners to test AI tutoring tools before they go anywhere near a classroom. The simulators are trained on anonymized learner records and model two specific things: whether the simulated student responds like a real one, and whether they update their answers after receiving guidance. Tested across chess, English writing, and math, StudentSim outperformed existing baselines on both dimensions.
Why it matters: The bottleneck for AI tutoring has always been feedback — you can't rapidly iterate on a tutoring system if every test requires real students. StudentSim makes that feedback synthetic and scalable. EdTech companies that get this working first will train better tutors faster than anyone running traditional studies. The research is not a product yet, but the tooling for whoever builds that product just got more serious.