Coding agents control robots through ROS 2

Following yesterday’s discussion of agent harnesses, OpenRUA gives Claude Code with Opus 5 terminal access to ROS 2 to write its own perception and control programs. The authors report 99% success on CaP-Bench and 87% on LIBERO-PRO in simulation, without task-specific training or bespoke robot primitives.

DeReAct separates action approval from completion checks

arXiv

Across GAIA and SWE-bench Verified, separate action and completion policies improve Sonnet 4.5’s first-attempt pass rate by 4.2–5.2 percentage points. With Opus 4.5, the pass rate stays comparable to ReAct, while execution records contain more complete evidence and better satisfy constraints.

Fast decision models need end-to-end cost accounting

arXiv

A paired evaluation finds that neither Laya nor Jev beats chance on zero-shot model routing. The authors’ audit also revises a claimed 23.9% cost saving to 4.3% after including a pre-screening cost.

Agents create and load their own tools at runtime

AI Engineer / AWS

This AWS talk shows agents writing and loading tools at runtime, then creating specialized sub-agents. The design starts with an editor, a shell, and a loader; the speaker emphasizes explicit token limits and evaluation.

Sourcegraph takes batch changes across repositories

AI Engineer / Sourcegraph

Sourcegraph describes a beta tool that rolls out changes across hundreds or thousands of repositories and responds to CI results and PR comments. In early Mercari testing, the company reports patching a GitHub code-injection vulnerability in 80 additional repositories beyond the initial targets.

AssemblyAI reports 80% support resolution for Joey

AI Engineer / AssemblyAI

AssemblyAI reports that its Joey support agent reached 80% end-to-end resolution within a week, at roughly $700 a month for tokens and inference. The agent uses synced local Markdown documentation and escalates the remaining inquiries to humans.

PayPal tests richer product catalogs for agent discovery

AI Engineer / PayPal

PayPal reports that targeted catalog enrichment improves agent recommendation accuracy, with the largest gains for products whose data is sparse. Excess unstructured text and marketing boilerplate can dilute semantic retrieval and reduce performance.

Reflection’s first open-weight model is expected this month

Axios via Techmeme

Axios sources say Reflection AI’s first open-weight model is due this month, alongside other Western releases. Its reported competitiveness with leading Chinese open-weight models remains a pre-release claim.

Most reference tokens match a small model

arXiv

A 500-million-parameter model reproduces 92–95% of reference tokens when given the correct preceding tokens. A routing experiment lowers projected latency from 7.59 to 5.12 seconds. The agreement rate measures token reproduction under that condition, rather than independent answer quality.

AI billionaires at Trump’s lunch spent $145 million on candidates

Forbes

Forbes tallies $145 million in political spending by AI billionaires who attended Trump’s lunch. The report examines differences in their spending despite a shared public position on AI regulation.

Musk says SpaceXAI will become SpaceXSI

Reuters via Techmeme

Elon Musk says SpaceX will rename its AI unit SpaceXSI, Reuters reports. This follows Trump’s push to replace artificial intelligence with super intelligence.

Altman describes a continuing split with Anthropic on regulation

Politico via Techmeme

Sam Altman says OpenAI and Anthropic still have fundamentally different views on AI regulation. He argues that AI’s benefits justify accepting some risks.

IWF reports rising AI abuse imagery

The Guardian

IWF assessed 6,310 AI-generated child sexual abuse images this year, exceeding the 4,500-plus recorded across 2025.