A single daily edition of announcements from the AI coding tools you use, read from a curated source list, summarised, and linked back to the original. No infinite feed, no account, no advertising.
Your pull requests and Slack threads can now start an agent without you.
Cursor's August 19 release lets cloud agents pick up work in response to events: an agent can monitor your pull requests, watch a Slack thread or run scheduled tasks rather than waiting to be started by hand. Subagents can now run on their own virtual machines, each with an isolated copy of the project and clean context in its own cloud environment, and /goal gives an agent a long-lived objective to work toward until it is fully complete. Any skill can be pinned in the chat as a Custom Mode, and a message sent to steer a working agent now waits for the next tool call instead of cutting it off mid-action.
Claude Code 2.1.237 fixes prompt caching for sessions running through an LLM gateway or a custom base URL, and adds a built-in Concise output style in which Claude leads with results and skips preamble and narration. The 2.1.236 release before it introduced ANTHROPIC_DEFAULT_MODEL, an environment variable setting the model new sessions start on, still overridable per session with /model. On macOS, wildcard read-deny rules such as a pattern matching every .env file now take precedence inside allowed regions, and a new notify_when_idle option lets one session ask another to report when it goes idle.
Turning reasoning off does not stop a model from reasoning where you can see it.
Armin Ronacher argues that a reasoning trace is a trained text behaviour rather than a system-level feature: a model learns to emit thinking into a designated channel marked by special tokens, and is trained to keep scratch work out of the final channel. Disrupt that convention and the tokens surface elsewhere — he notes that models with thinking disabled have reasoned into a bash tool and echoed the result to /dev/null, and that custom tools can trick a model into putting reasoning where it should not go. He also points out that reasoning effort is baked into the system prompt rather than passed as a sampling parameter, so changing the effort invalidates the KV cache.
Code your agent just wrote can run inside a hardware-isolated VM, not a container.
Simon Willison had Claude Fable 5 evaluate smolvm 1.8.3, a sandbox that isolates execution in hardware-backed VMs rather than shared-kernel containers, as a host for untrusted Python and JavaScript. The agent hit a wall in Claude Code for web, which has no nested virtualisation, and pivoted to GitHub Actions runners where /dev/kvm was available. The run confirmed no-network execution, read-only input mounts with writable output mounts, an unprivileged mode, guest-enforced timeouts against infinite loops, storage quotas and working CPU and RAM limits, at cold starts around 0.6 to 1.5 seconds and warm executions near 50 ms.
Your users could write the features you never shipped, if the sandbox holds.
Jeremy Morrell argues that LLMs radically lower the cost of authoring extensions, opening what he calls a new opportunity for extensible software on the web: applications with a stable core that users customise through generated code, serving the long tail a static product cannot. The safety half of the argument is capability-based security over ambient I/O, where code can only act through the references it has been passed, plus enforced limits on CPU, memory, network request count and size, response size and log volume, so that errors or infinite loops in a user's code never take down your service. He points at Salesforce, extensible at scale since 2007, as evidence the model need not trade away security.
