Agent autonomy expands: event-driven tasks, no-repo starts, and spend guardrails

Event-driven agents and frictionless starts

The biggest shift this week is toward agents that meet you where your work happens. Codex's event triggers let scheduled tasks fire from Gmail messages (filtered by sender or subject), Slack channel activity, or GitHub pull request events — so a review request can kick off a code change without you remembering to schedule it. Meanwhile, Cursor cloud agents no longer require a connected GitHub repo: pick "Start from scratch," prompt the agent, and it works in a temporary environment you can save to a Cursor Origin repo later. That's a meaningful lowering of the barrier for one-off automations and experiments.

The theme continues in Codex CLI 0.150.0, which adds task @-mentions so agents can read, create, or message other tasks from the terminal — useful for chaining multi-step workflows. The mobile counterpart, Codex for iOS, gains task search, a reasoning-effort gauge, and Home Screen shortcuts, making remote agent management more practical.

Tighter guardrails and spend visibility

Autonomy needs guardrails, and Claude Code shipped two notable ones. v2.1.248 adds --restricted mode, which strips out command/code execution tools and WebFetch unless explicitly named, keeps file tools inside the working directory, and refuses bypassPermissions — ideal for untrusted codebases or CI. v2.1.247 introduces /claude-api cost-optimize to profile and reduce API spend, plus a SendFeedback tool for drafting bug reports. And v2.1.251 adds a spend limit bar to /usage and a rate_limits.spend_limit status field — direct visibility into runaway costs.

Also worth noting: v2.1.246 fixes a credential leak where telemetry to Anthropic could carry a third-party gateway's API key, and v2.1.245 patches a startup crash on glibc 2.44 Linux distros like Arch.

Subagent infrastructure matures

The plumbing around agents improved across the board. Grok Build's 1.0.13 auto-continues truncated responses and retries transient inference failures, while 1.0.12 retries flaky MCP connections and fixes table-cell copying. Claude Code v2.1.251 streams live tool calls from foreground subagents to Remote Control clients — a big debugging win — and adds PreModelSwitch/PostModelSwitch hooks for intercepting model changes.

Devin's August 26 update improves Slack @mention routing and moves code scan pages under /security with batch remediation API routes. Codex CLI's 0.150.1 counts retained images toward the remote compaction token budget, preventing silent context loss. The consistent thread: agents are getting more reliable under failure, more observable in flight, and more respectful of your budget.