Every processed story in chronological order, with the newest coverage first. Filter by tag, source, or score to drill in.
The talk documents a concrete, production-tested eval architecture that closed the loop between offline simulation and live agent behavior at scale, directly enabling Lyft's resolution rate to climb from 10% to 35%.
The conversation grounds the limits of AI in science not in vague model capability gaps but in a concrete, structural problem: the physical world generates data too slowly and requires too much specialized tacit knowledge for AI reasoning alone to bypass it.
The construction removes the need for clearing houses and custodians in agent-to-agent forward trades by replacing institutional intermediaries with two HTLC contracts and one shared secret, making binding forward settlement possible between fully anonymous software counterparties.
SkillAudit removes the dependency on privileged external feedback signals that existing skill-evolution methods require, enabling agent skill improvement in real-world deployments where only a task description and workspace data are available.
AgentSpec provides the first controlled compositional foundation for studying embodied LLM agents, revealing that scaffold interaction effects — not individual module quality — determine performance, which reframes how agent systems should be designed and compared.
Heterogeneous model pairs using tap recorded defects or requested changes in 69.8% of reviews versus 53.1% for homogeneous pairs, demonstrating that cross-vendor agent collaboration on a shared codebase produces broader code review coverage than single-vendor setups.
The paper establishes a fundamental, mathematically proven ceiling on multi-agent system performance that is determined by task structure — specifically C_min — meaning agent scaling and increased communication cannot overcome poorly structured tasks.
The benchmark reveals that dialogue capability is a distinct dimension of coding agent performance not captured by existing autonomous-system evaluations, exposing a gap between how agents are benchmarked and how they are actually used.
The framework replaces the standard text-concatenation bottleneck in multi-agent synthesis with direct KV cache consumption, cutting time-to-first-token by up to 11x while preserving or improving task accuracy across diverse benchmarks.
GTBP directly addresses the two key failure modes of existing context adaptation methods — inaccurate credit assignment and lack of convergence guarantees — in multi-LLM agentic pipelines, providing both theoretical stability proofs and empirical gains across three benchmarks.