Google says an unattended 11-hour run of Gemini 4 Argon produced two accepted optimisations to its Agent Development Kit codebase. They cut in-memory session preparation overhead by about 40% in long sessions and cold package import time by 21%.
TechCrunch has compiled a list of notable AI agents that work through text messaging, ranging from general assistants to agents built for families, travel and work.
The latest OpenClaw release focuses on recovering interrupted agent sessions safely: they now keep track of where recovery came from and of unfinished tool results. It also keeps rejected Claude CLI drafts out of final replies and stops Codex previews from blocking answers after a context refresh.
OpenClaw 2026.9.34 is an extended-stable release, the project's equivalent of LTS, based on the late-September code with critical security updates, reliability fixes and support for new models. Its headline change makes apply_patch validate the whole set of edits before touching the workspace, so a bad later step can no longer delete files changed earlier.
Anthropic's open-source sandbox-runtime, the sandbox used by Claude Code, limits file and network access for agents, MCP servers and command-line tools. It relies on Seatbelt on macOS and bubblewrap on Linux rather than containers.
A tutorial shows how to build an engineering copilot with Letta whose memory goes beyond the prompt. It covers inspectable core memory, vector-based archival retrieval, state shared between agents, custom tools and portable workflows.