Across talks and results from AWS, Figma, Anthropic and Uber today, the reported constraint on agent work was never the model. AWS runs Lean proofs against Cedar's authorization semantics because neither tests nor human review can certify output for every input, and Figma's slowest adopters turned out to be its strongest engineers, the ones who see the failure modes first and want verification built before anyone runs an adoption campaign. Anthropic's automated alignment result points the same way: gains transferred to unseen benchmarks and to models 4.7 times larger, but only for failure modes somebody had already built a measure for. The practical read is that the next gain comes from the checking layer, and teams without one are buying throughput they cannot inspect.
An AWS engineer described an AI spending a week rewriting zlib in Lean, emitting 32,000 lines of proof decomposed into lemmas that a small independent kernel can check. AWS already runs this shape in production on Cedar, whose authorization semantics live in Lean while the shipping code is Rust, reconciled by roughly 100 million differential tests nightly. The argument is that a proof kernel scales to agent throughput in a way review does not.
Anthropic Fellows gave Claude 48 hours and a single GPU to research, train and test alignment methods on small models across ten failure modes including deception and sycophancy, with capabilities held constant. Gains held on held-out benchmarks and on models 4.7 times larger. The ceiling is the measure: the method only reaches failure modes somebody had already defined an eval for.
A dated brief from the vibe-coding frontier. Today’s Intel.
Figma's slowest agent adopters were the engineers holding the codebase together with undocumented context, which is exactly why they hit every failure mode first. Blum's read is to treat their objections as an ordered backlog of missing verification rather than as resistance, and to close that backlog before running any internal campaign. He also marks the human-written line at the top of an AI-generated PR description so reviewers know where to spend attention.
An OCaml maintainer saw path traversal probes within ten minutes of a patch discussion going public, then reproduced the exploit hunt himself using agents. rclone's maintainer reports 40 security disclosures last month against roughly 20 across the project's first decade. Embargo and staged release practice assumes a window between public hint and working exploit that no longer exists.
OpenAI shipped appshots, a keyboard shortcut that hands Codex and ChatGPT Work the full context of whatever application is on screen. GitHub added agentic Copilot runs triggered from Slack and Teams plus a multi session sidebar in Copilot CLI, and Claude Code desktop can now resume a session started in the terminal. The shared move is deleting the copy and paste between where the work lives and where the agent runs.
OpenAI is terminating the arrangement that gave Cursor direct access to its models, following Cursor's acquisition by SpaceX. Teams with GPT models wired into a Cursor workflow need either a different provider inside the editor or a different editor, and the timing is set by the contract wind down rather than by them.
DoorDash described a division of labor where annotation and eval definition sit outside engineering, held together by stable eval APIs rather than by trying to anticipate every annotation UI. Uber's uReview team named the metrics that made an AI code reviewer trustworthy: reply sentiment, whether a comment actually got addressed, and the agent's trajectory, because a model never signals that its own review is wrong.