Addy Osmani joins Claude Code

The stateless design flips the usual pattern: instead of another agent, Google ships a tool your current agent calls, keeping context ownership with Claude Code or Cursor. Developers building multi-tool agent pipelines can add Gemini's multimodal capabilities—video, speech, PDF analysis—without leaving their existing setup. It competes less with coding agents and more with direct SDK integrations, offering a uniform CLI interface any agent can invoke.

Philipp Schmid@_philschmid3 sourcesSource

The gains come from cutting agent steps and input tokens by more than half, which lowers API costs on retrieval-heavy workloads rather than just raising accuracy a few points. Teams building agents over enterprise document stores can plug it into Claude Code, Codex or Hermes Agent as a tool. The OfficeQA benchmark result offers a direct comparison point for evaluating retrieval layers before committing to one.

TestingCatalog@testingcatalog2 sourcesSource

The failure mode is not execution but the absence of a trigger to re-open strategy decisions mid-run, a finding that reshapes how agent benchmarks should be designed. Anyone building autonomous ML or post-training agents should plan checkpoints for human strategy input rather than longer budgets. Defaults also follow the harness, not the task — Claude Code picks full-parameter SFT 71.9% of the time while Codex CLI picks PEFT/LoRA 89.5% — so tool choice itself biases your pipeline.

OpenBMB@OpenBMBSource