It's 3 a.m. and production is down. What if the first responder wasn't you? Cognition just made Devin Oncall generally available. It picks up production alerts and bug reports before an engineer does, then: → Digs through logs and code → Escalates with the severity, the likely root cause and a suggested next step → Or opens a tested PR to fix it It also remembers past incidents. By recognising known issues and grouping related alerts, it skips about 60% of repeat investigations. The customer numbers Cognition shared: → AngelList: 50% of incidents handled end to end, and 3.5x more correct root causes than its in-house tool → Reducto: Devin flagged a problem from early warning signs and opened an incident before the production alerts fired The design choice which is most interesting: a second "review" agent runs alongside the investigator. Its job is to catch weak reasoning, like treating a recent deploy as the cause just because the timing lines up. That's the same pattern showing up everywhere this month. One agent does the work. A separate one checks it. On-call is a great place for agents, because the first hour of an incident is mostly gathering context. But it's better to start with "investigate and escalate" and earn trust before letting it ship fixes on its own. A good trial: replay last month's incidents and compare its root cause with what your team actually found.

Your AI agent can delete files, run up a huge API bill, or tell you "done" when it isn't. So who's supposed to stop it? Most off-the-shelf agents hide their limits in default settings your team never chose. A harness moves those decisions into code you own. ML researcher Elvis Saravia just published a walkthrough that builds one with the Pi SDK and the Jev decision model. It has 6 parts: Guard the tools: check every call before it runs Set rules you can change: allow, block, or ask a person Route by difficulty: a cheap model for easy work, a strong one for hard work Plan for silence: if the checker doesn't respond, block tool calls Check the answer: compare the reply with the files the agent actually read Keep a record: log every decision and the numbers behind it The result: your agent's limits, model choices and human handoffs are all readable in code. This is what client work needs. Clients don't trust "the AI is smart." They trust limits they can see. The full breakdown, with an example rule function, is in the carousel. Which of the 6 is your agent missing right now?

Every LLM you use writes one token at a time. This one doesn't. Inception just launched Mercury 3, a diffusion LLM. Instead of writing left to right, it drafts the whole answer in parallel and then refines it, a bit like an image model sharpening a picture. The numbers: → 1,068 tokens per second on Nvidia GPUs → 1M-token context window → $0.10 per million input tokens at launch (50% off) → OpenAI-compatible, so it's close to a drop-in swap Inception claims it beats GPT-6 Luna and Claude Haiku 5.5 on agentic coding. Haiku still wins on longer coding tasks. These are vendor numbers, so test before you trust them. Why speed matters this much: agents wait on the model at every step. In a voice call, a browser agent or a tutoring app, a few seconds per step adds up to an experience that feels broken. For real-time products, "smart enough and fast" often beats "smartest". Where would 1,000 tokens per second change your product?

Two new voice-agent stacks shipped on the same day. They take opposite approaches, and the difference matters if you're building one. ElevenLabs — Eleven Realtime One speech-to-speech model hears and speaks. → Max starts talking in 0.57 s and scores 98.8% on Big Bench Audio → Mini is the cheaper option for natural chat. It handles interruptions and "mm-hmm" → Already on OpenAI's Realtime API? Change the endpoint and model name Mistral — Voxtral 3 A pipeline: listen → reason → speak, streamed together. → Each turn finishes in under a second → Swap the LLM, voice or language without rewriting the agent → Runs on Mistral's EU-hosted API from $0.04/min, or on your own servers For building AI voice receptionists, here's how to choose: Speech-to-speech when a fast, natural feel is what sells. Pipeline when the client needs control: their own LLM, their own data region, a specific language. Either way, the voice itself is no longer the bottleneck. The tool calls and the script are. Which would you pick for building AI voice receptionist?

Banks are finding AI agents inside customer accounts that nobody at the bank authorized. According to Axios, major U.S. banks are detecting third-party agents in customer accounts, and some are taking actions the customer never asked for. Executives say they struggle to tell a human from an agent. Most of these agents got in the simplest way possible: customers shared their own logins. Same week, Microsoft said China-linked attackers used an open-source AI security tool to break into IT firm ConnectWise, then tried to log into 2,250+ customer environments. At some, they had high-level access within three hours. The fix isn't banning agents. It's never giving them the password. 1Password just announced 1Password for Agents: 1- The agent reaches a sign-in form 2- It requests a credential from your vault 3- You approve 4- 1Password fills it in on the site. The agent never sees it. It works with Claude, OpenAI, Grok Bot, Meta's Muse, Perplexity, Browserbase, Browser Use and more. It's free and opt-in for personal and family accounts. If you build agents for clients, copy this pattern: give the agent access to do the task, never possession of the secret. Would you let an agent log into your bank, even with an approval step?