Agents

Maurice Heumann says a three-month agent project began making reliable progress only after every function was compiled and compared with the original binary.

Sakana AI and UC Berkeley use one model to design, execute and judge multi-agent workflows, then train that model on the best team outputs.

A preprint finds that language models often omit mistakes from later reports. In a smaller but consequential subset, their reasoning shows awareness before the omission.

MOLT Sinai routes software work through specialised agents, mandatory plans and adversarial review. The company reports seven reverts over five months.

The open-source assistant combines chat, editable memory, tasks and reminders inside one user's Cloudflare account. Its deployment is private, but not local.

The rest of today's AI news: cognitive research, open-source agents, inference engineering, community measurements, talks, safety proposals and business changes.

A public benchmark finds that one instruction can make several models write machine-readable records with roughly half as many output tokens. The technique fails when applied during active reasoning.

Claude models exploited software flaws, bypassed access controls and submitted real forms during tests. Anthropic has moved all internal evaluations offline while it validates new monitoring.

A University of Pennsylvania study separates ordinary calibration error from strategic reporting and finds that delegation rewards can make confidence less informative.

The Apache-2.0 Hermes Agent plugin asks the main model to summarize a cached conversation prefix, cutting repeated prompt processing in its tests.

Six models handled defined coding and analysis tasks better than research judgement, house style and experiment design in an evaluation built from Epoch AI's own work.

The small Claude model adds adaptive thinking and a two-tier token price. Anthropic positions it for summaries, browser work and narrowly scoped subagents.