Block’s human-sounding voice AI runs on laptops

Welcome back. Moonshot just dropped Kimi K3.1 Pro, and its coding performance claims and affordability look like it could give our go-to AI model some serious competition. Meanwhile, a new voice model is here, and we can’t get over how eerily human it sounds.

Also: A senior dev reveals the new go-to language for AI agents, a veteran engineer shares his daily coding workflow, and a clever Claude Code trick to troubleshoot your Mac. 

Block’s human-sounding voice AI runs on laptops: Jack Dorsey’s fintech company just unveiled Dessa, a full-duplex model trained entirely on human speech. Unlike most voice systems, it doesn’t need an external LLM and works directly with audio, picking up sarcasm, laughter, and interruptions along the way. Block says it can run locally on a MacBook Pro M4 Max or better. Access is limited to safety researchers for now, but you can hear how human it sounds in the demos.

Moonshot’s new open model closes in on frontier coding: The Chinese AI startup just dropped Kimi K3.1 Pro, a 1M-token context model that scored 65.7% on Terminal-Bench 4.0, per the company. That puts it ahead of GPT-6 Astra and just behind Claude Opus 5.5, at $2/$10 per million input/output tokens. It handles long-horizon coding tasks, from quick bug fixes to building games in Unity and Godot. Moonshot also deliberately limited its offensive cyber skills by removing hacking-related data from the training set. Try it here, with open weights arriving by Nov. 4.

Tailscale SSH flaws could hand hackers root access across servers: If your team relies on the mesh VPN for SSH, you will want to pay attention to this. Researcher Erica Windisch just published 11 flaws in its software and the self-hosted headscale server. In her lab, one hacked device abused SSH port forwarding to take full control of other machines and add a rogue device to the network. Tailscale says fixes are coming, but there’s no patch yet. For now, audit SSH access and disable unnecessary forwarding.

Apps and agents increasingly need to run long-running background tasks but serverless is too restrictive, and managing workers and queues by hand is a headache. Render Workflows gives you the best of both in a lightweight SDK.

With Render Workflows, you can:

Try it today and get $50 in credits to run your first workflow.

You know the drill. Someone asks for a ship date before you even know what is lurking in the code. In complex systems, research eats up most of your time, making accurate estimates impossible. GitHub staff engineer Sean Goedecke agrees with this and in an essay he shows how he still finds a way to give an estimate anyway.

He starts with the date. To Goedecke, estimates are tools for deciding which projects get funded or cut. Managers often have a timeline in mind before the design exists, so he works backwards from it.

His process looks like this:

The real skill is earning trust. Your manager picks the tradeoff. Goedecke saves “this is impossible” for the rare cases where it really is, because people stop listening to someone who pushes back on every estimate.

There is one big catch. Companies often treat estimates as if they transfer between teams. In Goedecke’s experience, some teams ship the same work ten times faster than others, so a number means little until it’s clear who is doing the work. If you’re ready to go deeper, check out Goedecke's piece, "Engineers Who Won't Commit," to learn how to make the call under uncertainty. 

Scrape complex targets without the constant maintenance. Decodo’s Web Data API rotates proxies and executes page scripts automatically to give you structured web data.

Claim 10K Requests

How to run your AI software factory like a real engineering org: This tutorial breaks down how to scale a solo developer's AI agent workflow across an entire team. It walks through assigning clear ownership and standards to agents, setting up skill and agent registries, using readiness scorecards, and keeping humans in the loop for critical sign-offs. 

HumanLayer (12.5k ⭐): Run Claude Code, Codex, OpenCode, and Pi agents across your own machines, then manage all their tasks, sessions, plans, and reviews from one place. It’s built for teams that want multiple coding agents working in parallel without losing track of what each one is doing.

How to simplify your multi-agent setup (by Anthropic): This guide argues you probably need fewer subagents than you think. It shows when to fold planners, routers, and specialist agents back into one main loop using skills and tools, and when separate subagents still make sense for parallel work, fresh context, or overloaded toolsets.

A senior dev got tired of explaining AI’s most misunderstood term, so he wrote the definitive guide to working with harnesses.

If you enjoyed this newsletter, you should definitely check out our IT and cybersecurity newsletters. Subscribe using the links below:

We put together 20+ hands-on guides covering coding agents, agentic engineering, and leadership frameworks to help you level up your career. You can get them here.

Advertise with us: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team