When the scores match but the work does not
Behind success rates in the eighties, 558 trajectories reveal different habits of reasoning and tool use
Core topic
AI agents
Moving AI agents from tools to accountable collaborators
Agent orchestration, collaboration, ownership, escalation, long-running work, and human oversight from an operational perspective.
Related stories
Page 2
Whose doubt do you trust when everyone says yes, CUSP and collective uncertainty
CUSP pools answers from several VLMs in a shared semantic space and splits total dispersion from model disagreement, and this piece follows the design and the numbers to rethink trust in multi-agent reasoning.
Bringing agentic evaluation into one harbor: Harbor Adapters and Harbor-Index
A research commentary on unified infrastructure across more than eighty agentic benchmarks and a curated meta-dataset of eighty two tasks
Iris: How to Climb to the Search Frontier
Tracing how Iris-mini and Iris-pro reverse-build questions from hyperlinks, mask clues as descriptions, and alternate SFT with live-search reinforcement learning to reach frontier benchmark scores.
Rethinking indirect prompt injection as test-time search
A research commentary on treating attack success as a function of attacker search budget and procedure rather than a fixed victim score
When Better Models Make Riskier Systems
Why stronger models can raise system level risk, explained through LLM agents in financial markets with an engineering lens.
Reading research acceleration from inside OpenAI
Following Simon Willison's notes on OpenAI's internal research acceleration, a walk through the RSI wording, daily coding agent use, and the sharp bend in AI spend per researcher.
HarnessDev: Can LLMs create and evolve their own agent harness?
Instead of hand built agent frameworks, a loop where a language model writes harness code and keeps the variants that perform. A close read of the idea, the mechanism, and the hard parts.
Reading Repo-To-Skill: what agents lack is not models but operational knowledge
Following the DisCo paper on distilling operational knowledge into skills, with a close look at the AREX-Skill library and what the matched-budget experiments actually show.