Slack native coding agents are here

Welcome back. Last month, Twitter cofounder Jack Dorsey dropped Buzz — a workspace where agents and devs can work in sync. It looked like the first real threat to Slack's dominance. Now Slack is punching back. Instead of forcing agents to live in isolated terminals, they’re bringing them straight into your daily workspace to redefine how you ship code. More details below.

Also: Claude Code’s new output setting cuts the small talk, a simple solution for draining batteries of a MacBook, and when to build your own agent harness.

Slack native coding agents are here: Salesforce just launched Slack Code, which pulls AI agents out of the terminal and puts them right where your team already hangs out. Now, you can spin up an agent like Claude directly in a Slack thread, give it context from your conversation, and watch it code and preview progress in real time. Instead of one developer working with an agent alone, coding becomes multiplayer. Watch how it works.

YC-backed startup ships VS Code inside your terminal: Zenbu Labs just open-sourced terminal-code. So you can type one short command, and a full VS Code pops up right in your terminal. Devs can also get diffs, split panes, and extensions, plus their settings imported from other editors. It even matches your color theme. For now, it runs on macOS and Linux. Windows users can access it through WSL (Windows Subsystem for Linux), a system which lets Windows 10 and 11 run Linux directly without a separate virtual machine or dual-boot setup. Try it here.

Ramp takes on rising model costs: The fintech startup just dropped Router.com. A single endpoint that automatically routes every request to the cheapest model that still hits your quality bar. Since the API is compatible with both OpenAI and Anthropic, teams can switch providers with just one line of code. Ramp says Router cuts AI costs by 40% on average, and can automatically fall back to another model if a provider goes down. It’s free through 2026, with $26 in credits to start. More details on the API here.

Most AI models start from zero, with no memory of your tone, voice, or cadence. Essentially, they just mimic the average writer on the internet, which is downright boring. 

In Oracle’s latest step-by-step tutorial, you can learn how to create three layers of agent memory with Oracle AI Database, so agents can write engaging social posts in your voice. 

Read the agent memory playbook here.

The margins. AI features look attractive until you're scaling to millions of users. Canva learned this the hard way. After nine straight years of profitability, the company slashed its 2026 growth forecast from 30% to 20%. CEO Melanie Perkins attributed the shift to the massive overhead of serving AI to their 265 million monthly users.

Hit the brakes. So Canva paused its AI rollout to focus on the basics. They audited their systems to see what was actually happening and find ways to improve. Here's what they did:

The playbook is simple. These changes reportedly slashed serving costs by about 90%. The same strategy works for smaller teams: start lean and scale only when necessary. Offload simple tasks to smaller models, cache repeated context, tighten output lengths, and consider self-hosting once your volume justifies the overhead.

PS: Find the full discussion here. For routing specifically, RouteLLM drops straight into the OpenAI client and shifts simpler queries to cheaper models automatically.

The intelligence era is here. Join Glean:GO 2026, Aug. 26–27 to learn how enterprises turn AI into measurable impact. Hear from leaders from OpenAI, NVIDIA, and Snowflake, and gain practical strategies from keynotes and hands-on sessions.

Register now.

When to build a custom agent harness (by LangChain CEO): Check out this deep dive from LangChain co-founder Harrison Chase. He walks through exactly how to build your own custom AI agent harnesses. You’ll learn to create custom middleware, implement observability, and establish data flywheels to optimize production-ready agents.

Anthropic Cybersecurity Skills (30.1k⭐️): This tool gives AI agents 817 ready-made security skills across 29 domains, with support for Claude Code, Codex CLI, Cursor and more.

How to build on the open agent harness (by OpenAI): Coding agents are powerful, but jumping between tools is a hassle. This guide shows you how to use the open Codex harness to bring agents directly into your workflow, complete with context, tools, and approvals. Using this approach, Asana finished a massive frontend migration in just two weeks, a project that was originally expected to take five years. 

Codex normally stops after one turn. For a long task, such as a migration or a CI fix, use a persistent goal.

Add a clear completion rule:

Control the goal:

If “/goal” is not in the slash command list, enable goals:

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

Ever wondered which AI models your computer can actually handle? An open-source tool profiles your hardware to find out for you.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 300K+ 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