AI Agents

An observational study of tens of thousands of Microsoft engineers during the early-2026 CLI-agent rollout. First use spread through social networks, retention correlated with coding activity rather than tenure or role, and adopters merged roughly 24% more PRs over a four-month window with the lift persisting. The authors are candid that merged PRs proxy output, not delivered value. Still, this is a rare large-N field measurement rather than a lab benchmark, and the peer-diffusion finding is directly actionable for rollout planning.

Archer targets LLVM optimization review, constraining the agent from both ends: obligations guide the analysis, and a deterministic validation guard admits only findings backed by executable evidence. On 70 open and 328 closed recent PRs it flagged 21% and 11% respectively as introducing semantic bugs like miscompilations. The evidence-gating is the interesting part—it converts a chatty reviewer into one that only reports what it can demonstrate. If the flagged rate holds under expert triage, it exposes a real review-capacity gap in a critical codebase.

As planning graphs grow to thousands of steps, full-graph replanning after each failure burns context and degrades long-context retrieval. WM-SAR works backward from subgraph amplification to isolate the nodes and edges that keep re-amplifying error, sending only that causal subgraph to the LLM for in-place repair. Against symptom-scanning engineering correctors, it achieves near-whole-graph stabilization from a compact region under realistic token budgets. The framing—repair the amplifier, not the visible symptom—is a cleaner mental model for correcting persistent workflows than local patching.

SkillCoach argues final-verifier success is too coarse when skills overlap: an agent can pass by trial-and-error while picking distractor skills, skipping steps, or omitting checks. It derives skill-grounded process rubrics from real rollouts and scores trajectories along four axes—selection, following, composition, and reflection—keeping the external verifier as a separate outcome signal. The evolved rubrics both surface failures hidden by accuracy and act as process supervision for filtering training trajectories, reportedly beating outcome-only filtering. Separating process quality from accidental success is the right instinct for skill-heavy agents.

A game-theoretic treatment of runtime human oversight where information runs both ways: the human privately knows her reward, the AI privately knows its proposed action's quality. Built on CIRL and the Oversight Game, it uses a contextual-bandit structure to get exact one-shot characterizations impossible in the full POMDP. The result: a 'slab of avoidable harm' where the AI knows an action is harmful and shutdown would help, but a myopic human trusting her prior declines to oversee. That gap is framed as the price of non-credible oversight communication.