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.
The Codex harness comes as a managed API, sandboxes and subagents included.
OpenAI opened the Agents API in public beta on 10 September, giving applications the same managed harness that runs Codex: sandboxed command and code execution, tool and MCP connections, context summarisation, subagent delegation, and session resumption and steering mid-run. Compute runs either in an OpenAI-hosted sandbox that handles sessions and recovery, or on your own self-hosted infrastructure, configured with a workspace directory and capability directories. Billing adds nothing beyond standard model, tool and container rates, but the beta is restricted to US data residency and offers no zero-data-retention support even when the sandbox is your own.
SWE-2 buys most of Fable 5.1's coding score far cheaper, and folds on harder agentic work.
Cognition released SWE-2 on 10 September, post-trained from Moonshot's Kimi K3, a 2.8T-parameter base that had already been reinforcement-trained for agentic coding. It scores 50.0% on FrontierCode 1.1 Main against Fable 5.1's 50.9% and GPT-6 Astra's 53.3%, while being 64% cheaper, and its medium effort level makes a first real edit after a median of 18 steps where SWE-1.7 took 48. On Terminal-Bench 4 the gap reopens to 27.3% against Fable 5.1's 55.8%, and the model ships only inside Devin Desktop, the Devin CLI, Devin Web and Fusion, with no standalone API, pricing sheet or weights.
A Cursor project keeps one body of work in context for months.
Cursor launched Projects in beta on 10 September, rolling out to all users the same day. A Project takes on a feature, a migration or a full app, maintains context over months of work, and is run by a coordinator agent that writes no code itself: it plans the work, delegates it to implementation agents, and brings the finished work back for you to check, running as many in parallel as the work needs. Each Project syncs a file store across every machine its agents use, so research, artifacts and learned preferences about the codebase build up, and subscriptions let the coordinator watch a Slack channel, run on a schedule or follow all your PRs and act on what it detects without waiting for a prompt.
Claude Code's deny rules were not holding against env -C and eval.
Claude Code 2.1.268, released 10 September, fixes a case where a Read or Edit deny rule did not apply when an env -C, eval or similar command the permission checker cannot analyze was on the same line, and deny and ask rules on symlinked directories not applying when a path was given by its real location. The same release stops plugin and marketplace errors showing a token or password from a git source URL, and stops MCP server details and login errors showing secrets resolved from ${VAR} placeholders in MCP configs. WebFetch, which could hang indefinitely on a server that keeps the response open, now fails after 300 seconds, with CLAUDE_CODE_WEBFETCH_DEADLINE_MS to override the deadline.
Agent tasks in VS Code can run hourly, daily or weekly with nobody watching.
Visual Studio Code 1.137 shipped on 9 September with Agent Automations in preview, which schedule recurring agent tasks hourly, daily or weekly or run them on demand, from templates for catching up on recent changes, triaging issues and finding bugs, or from a custom prompt. A new Agent Host runs agent harnesses in a dedicated process and connects several windows to one session over the Agent Host Protocol, so a session survives its folder closing and can run remotely over SSH. Voice Mode ships as experimental, and the features are gated behind the chat.automations.enabled and agents.voice.enabled settings.
Autonomous exploit-writing loops turned up more than a dozen zero-days in one month.
Anthropic's September threat intelligence report says a majority of the operations it documents were enabled by AI through direct execution or orchestration, with multi-agent frameworks running reconnaissance, exploitation and data exfiltration. One Chinese espionage actor ran autonomous workflows that tasked the writing of exploit code iteratively until it worked, yielding more than a dozen possible zero-day findings in a single month, while a Russian espionage actor had Claude identify, modify and redeploy malware artifacts once defenders detected them. The AI supply chain became a target in its own right: actors stole AI API keys from victim environments and spent them as compute, and one election-manipulation operation used Claude Code to build dashboards for running networks of fake accounts.
Two native codebases no longer cost twice the work, so Shopify dropped React Native.
Shopify is moving its mobile apps off React Native and back to separate Swift and Kotlin codebases, reversing the decision it made in 2020. It says LLMs changed one of the core assumptions behind that decision, so it re-evaluated the stack from first principles: agents can implement a feature on Android using the iOS version as a reference and the other way round, cutting the cost of holding parity between platforms through shared specifications, tests and review checkpoints. The two-platform cost has not disappeared, but agents now do enough of the implementation, translation, testing and review that it is no longer the deciding factor it was in 2020.
