Tien geselecteerde signalen over agent-evals, wetenschappelijke benchmarks, multi-agent coherentie, AI-werkflows en domeinmodellen.
arXiv reasoning / agents / evals
Gram introduces automated alignment auditing for sabotage propensity in agentic coding and research deployments. This matters because serious agent rollout needs behavior-level evaluation, not just output review.
arXiv reasoning / agents / evals
ProjectionBench evaluates scientific hypothesis generation as information is progressively revealed. This is useful because real discovery work is uncertain and incremental, not a static benchmark lookup.
MIT Technology Review AI
MIT Technology Review covers the operational difficulty of controlling a Bundibugyo virus outbreak. The selected angle is operational readiness: detection, coordination and feedback loops matter more than isolated capability.
arXiv reasoning / agents / evals
This paper formalizes how multi-component LLM agents can be locally coherent but globally incoherent. It is directly relevant to workflow design, runtime checks and repair mechanisms for composed agent systems.
TechCrunch AI
TechCrunch discusses Aaron Levie's warning that executives deciding AI can replace jobs often understand those jobs poorly. The useful signal is the need for workflow understanding before automation claims.
TechCrunch AI
This podcast covers the same critique from Aaron Levie: AI replacement narratives often skip the actual work analysis. It supports a problem-first implementation position.
The Decoder
The Decoder summarizes a review paper arguing that tools, memory, tests and permission boundaries are the layer that turns a model into an agent. Useful frame: model plus harness equals agent.
TechCrunch AI
Cognition positions Devin as a coding agent that works with human programmers rather than replacing them. This points to delegation, review and workflow ownership as the mature agent pattern.
The Decoder
OpenAI is offering GPT-Rosalind through the Rosalind Biodefense program. The practical signal is domain AI as public infrastructure, where governance and evaluation matter as much as capability.
Google News AI Adoption
Google News surfaced an MSN item about PATH's agentic AI solutions and enterprise adoption. It is the weakest selected signal, but it shows the current market language Bart should pressure-test: what work, permissions and accountability actually change?