Vercel AI Gateway Adds Session View Tracking Costs and Cache Hits for Claude Code, Codex, CursorAmazon Uses AI to Monitor Union Activity at UK Warehouses in Birmingham, CoventryRealtimeBench Tests LLM Agents on a Live Clock With 180 Time-Sensitive TasksSelf-Improving Agents Keep Wrong Skills; Oracle-Free Operator Induction Targets Exact AnswersEvolving Together, Failing Apart: The Skill-Policy Co-evolution Trap in Self-Improving AgentsConfidence-Calibrated Inference Expansion Frames Reasoning as a Three-Action Control ProblemWhen Reasoning Traces Conflict With Evidence: Answer Configuration Shapes Evidence FollowingText-to-Video Models Miss Ordering; Formal Temporal Specifications Guide GenerationVercel AI Gateway Adds Session View Tracking Costs and Cache Hits for Claude Code, Codex, Cursor

Amazon has deployed AI systems to monitor union activity at its UK warehouses, according to Open Democracy, covering facilities in Birmingham, Coventry, Atlas, and Sutton Coldfield. The report says the system is used to detect signs of union organizing. The story drew 145 points and 57 comments on Hacker News, with discussion focused on the limits of workplace surveillance and employees’ right to organize. For UK warehouse workers, it means union organizing faces an added layer of algorithmic scrutiny; for AI governance watchers, it is a concrete case of workplace AI extending from productivity management into labor relations.

LLM agents can solve complex interactive tasks, but reasoning and tool calls take time, so in continuously changing environments an otherwise correct action may arrive too late. The authors introduce RealtimeBench, with 180 tasks across six categories, to evaluate agents under a live clock. It fills a gap most existing benchmarks ignore: correctness depends not just on what an agent does but on when it does it. For teams deploying agents, this means evaluation needs to treat latency and timeliness as first-class metrics, since strong offline scores may not translate into usable online behavior.

The paper identifies a practical failure in self-improving agents: an agent must decide which new skills to keep, and in practice it keeps whatever earns high reward, convinces a judge model, or compresses its skill library, even when the skill is wrong. That breaks down when answers must be exactly right. The authors propose oracle-free operator induction, in which the agent grows its own domain-specific language (DSL) to support exact question answering rather than relying on external judges or reward signals to filter skills. The takeaway: skill retention criteria must align with answer exactness, or self-improvement will reliably accumulate errors.

The paper examines skill-policy co-evolution in self-improving agents, where an agent’s policy model and its skill bank improve each other during training. The authors argue current evaluations have a blind spot: they score the co-evolved policy only on the final skill bank it trained with, so problems introduced by the co-evolution process itself go unseen. They call this the skill-policy co-evolution trap. The implication is a challenge to evaluation practice: scoring only the end state can mask degradation or mismatch during training, leading to overestimated robustness in deployment.

The paper frames evaluator-guided reasoning as a local control problem at every step: trust the current answer, return a fallback, or spend more inference. The authors formulate it as a compute-aware three-action controller that scores terminal actions by predicted values and uses confidence calibration to decide when to expand inference. The benefit is turning inference budget allocation from a fixed policy into a per-step dynamic decision. For cost-sensitive deployments, this kind of controller offers an explicit framework for trading accuracy against compute spend.

The paper notes that a reasoning trace retained at answer time plays two roles: it can raise accuracy, and it can compete with current evidence when the two disagree. By crossing verified evidence versions with traces generated from each, the authors disentangle these roles. The result points to an engineering question: reasoning traces are not unconditionally beneficial, and how the answer is configured, including whether and which trace is attached, changes how closely a model follows evidence. For evidence-grounded QA systems, trace retention policy needs to be designed per use case.

Text-to-video models often generate the visual concepts requested in natural language prompts but fail to satisfy their intended ordering, timing, and persistence requirements. The authors propose representing these requirements as specifications in a formal temporal logic and using those specifications to guide generation, rather than relying on natural language prompts alone. The work targets a known weakness in narrative coherence and action timing. For video generation developers, it suggests controllability can shift from prompt engineering toward verifiable formal constraints.