AI Agents

Agentic workflows release each ready turn immediately, and under contention that eager policy piles up submitted-but-unfinished work that can no longer be reordered, blowing out tail latency. This paper decouples readiness from release: a mean-CVaR objective decides which turn to release and how much in-flight work to tolerate, using online estimates of turn work and adapting the budget to queue pressure. On real SWE agent traces it matches eager release under light load and cuts P95 flow time up to 3.5x under contention.

Compound systems usually add a generative meta-agent to read worker outputs, assemble answers, and decide when to stop — concentrating three control decisions in one opaque, order-sensitive call. UnitBoost replaces it with a defined operator: a unit map turns outputs into slot proposals, a constrained argmax assembles the result, and unfilled slots become an explicit residual for the next round. It's order-free, records provenance, and beats input-matched generative managers by 0.048-0.076 across three benchmarks. Honest about limits — it names conditions where no gain exists.

Post-task memory curators only see completed trajectories, so they preserve errors, overgeneralize, and retain stale facts. Environment-probing curation gives the curator least-privilege, read-only world tools to verify and refresh candidate memories before writing — no retraining, no changes to the task agent or write path. In a GitHub Copilot harness on CLBench, probing lifts pass rate from 39% to 73% while cutting task-agent cost from $3.38 to $1.68. Gains hold across 90 APEX tasks and two model versions without schema drift.

Procedural agents fail invisibly: the final answer looks fine even when the agent skipped a required check, took the wrong branch, or violated an invariant. Output-only judging sees the answer, trace judging sees activity, but neither knows which obligations were active for the query. ContractEval represents instructions as query-active obligations and matches them against evidence, turning omissions, ordering errors, and invariant breaches into distinct failures. With gold graphs it localizes all injected structural failures; LLM-based extraction keeps much of the signal but is calibration-sensitive.

With thousands of tools available, agents get shown a short ordered menu and can call only what's on it. Relevance-ranked menus surface the final action but omit or delay the less obvious producers that create its inputs. This work treats the menu as an execution prior: a state path routes from the observable request to the outcome, the retriever covers producers plus the final action, and a reranker places producers before consumers. On ToolBench, online success rises from 0.737 to 0.898, and 32 well-chosen tools beat the official 128. Code released.

Amazon Science probes why ML research agents that iterate against a validation metric don't overfit it the way a naive optimizer would. The surprising-result framing is the draw: agents that repeatedly tune toward a score should game it, yet apparently don't collapse into the expected pathology. Worth reading for the mechanism they propose, though the HN discussion suggests the conclusion is more nuanced than "agents are safe." Treat the finding as scoped to their setup, not a general guarantee.

Amazon Science asks whether agreement between LLM judges is evidence of correctness or just shared bias. The distinction matters directly for anyone using ensembles of judges as a proxy for ground truth: if two models fail the same way, their consensus is worthless, not reassuring. The piece is short on methodological detail from the excerpt, but the core caution is right — correlated errors inflate confidence in eval pipelines that treat judge agreement as a quality signal.