AWS blog: giving agents more autonomy without losing control
Read original source →aws.amazon.com
AWS's Shawn Bice writes that enterprise leaders want agents to do more but don't trust them to do it right all the time, noting the same creativity and autonomy that makes agents useful makes them hard to trust.
1 source
More stories today
GPT-Live-1.1 updates Voice in ChatGPT, Codex, and dots
OpenAI·1 hour ago
Fink: Data Center Moratorium Would Tip US Into Recession
BlackRock CEO Larry Fink said a US moratorium on data center construction would trigger a recession, arguing "so much of our economy is based on the build-out of this capex." He called the backlash over power and water use "misinformation," as at least 100 localities approved moratoriums by July.
Decrypt·1 hour ago
Vercel makes Python generally available as first-class runtime
Vercel now runs CPython on Linux with automatic support for FastAPI, Flask, and Django, plus uv and pyproject.toml for dependency management. Beta Python SDKs cover AI SDK, Workflow, Queues, Sandbox, Connect, and Blob.
Vercel Blog·1 hour ago
Vercel launches Token Pass: $30/mo for $80 in GLM 5.3 and Kimi K3 credit
Token Pass is in early access for Pro teams, bundling $80 in monthly model credit for $30. It runs through AI Gateway with Zero Data Retention by default, and unused credit does not roll over.
Vercel Drives gives agents persistent storage across sandboxes
Vercel Drives is now generally available for Vercel Sandbox, letting agents mount the same drive in a new environment and resume work where they left off. Drives can be shared read-only across sandboxes or forked so each agent gets its own writable workspace.
Vercel expands enterprise AI governance platform
Vercel made Bring Your Own Cloud on AWS generally available and launched Enterprise-Managed Authorization via Vercel Connect in public beta. Enterprise token usage on Vercel's AI Gateway quadrupled in three months.
Devin Oncall launches to auto-investigate production incidents
Devin Oncall is generally available and scored 100% end-to-end success on SREGym, saturating the benchmark with the fastest time to mitigation. AngelList said Devin found root causes 3.5x more often than its in-house harness Sky and handled 50% of incidents end-to-end.
devin.ai·1 hour ago
Apple MLX runs trillion-parameter Muse Spark Open on four Mac Studios
Apple's new MLX release runs Muse Spark Open (816B params, NVFP4) at over 36 tokens/sec across a four-Mac Studio cluster via RDMA over Thunderbolt. Benchmarks compare M5 Ultra vs M3 Ultra on Muse Spark Open, Mistral Medium 3.5 (128B), and Qwen 3.8 (27B).
Apple ML Research·1 hour ago
