Anthropic's new model nails bug fixes

Welcome back. Opus 5.5 already set a high bar for how fast AI models can be, but Anthropic just raised it again with Sonnet 5.5. Meanwhile, Nvidia is stepping in to keep rogue agents from blowing past guardrails. 

Also: Why OpenAI pulled the GPT Astra 6.1 launch, a SpaceXAI engineer celebrates the death of coding, and a Meta engineer's roadmap to AI engineering.

Anthropic's new model excels at debugging: The AI lab just released Sonnet 5.5, a cheaper sibling to Opus 5.5. It runs up to 30% faster than Sonnet 5 and costs up to 30% less on most tasks. A senior Anthropic engineer shows how they compare when it comes to codebase debugging. Meanwhile, Claude Code also gets two new commands. “/claude-api build-eval” turns real traffic and bug reports into tests, while “/claude-api hillclimb” improves results one edit at a time, rolling back changes that overfit.

Nvidia rolls out new defenses against rogue AI agents: The chipmaker just shipped an Open Agent Safety Platform that lets engineering teams box in autonomous agents. OpenShell puts enforceable limits around what an agent can reach, outside the model itself, so you can control its access. If it still misbehaves, a companion watchdog called Sentry can quarantine it in milliseconds. Hugging Face has already wired it up to catch misuse. Watch it work. 

This new AI model is meant to think less: Fireworks Research just dropped Ember-1, a specialized model built on Kimi K3. It was tuned to think less, so it hits the same benchmark scores as K3 while using about 40% fewer tokens. Available in the Cline desktop app, that leaner reasoning means faster replies. It also points to a trend where inference providers are reworking open-weight models to make agentic coding cheaper. Try the beta.

Attackers are using AI to find exploits faster than ever. Are you still waiting for bi-annual pentests? SonarQube Hunter Agent puts an automated AI security researcher in your CI/CD pipeline.

It hunts for complex business logic flaws and broken access controls that traditional tools miss. There is no generic LLM prompt variability here; Hunter Agent is playbook-driven, repeatable, and verifies every flaw. Match machine-speed threats with machine-speed defense.

Put Hunter Agent to the Test

The big question. At a WorkOS event in New York, a developer asked veteran engineer Gergely Orosz the question on everyone's mind: "Is my job going away?" When Orosz posed that same question to engineers at a frontier AI lab, they could not agree on an answer either. While he could not offer a definitive answer, he shared four practical tips on what to do next: 

The next step for us. If you feel like you're falling behind with AI agents, you're not alone. We publish practical, thoroughly researched guides every Saturday to help you get up to speed with AI. The first step is mastering coding agents so you can orchestrate them effortlessly. You can get started with our Ultimate Agentic Engineering Guide right here, and be sure to check your inbox every Saturday for new editions. 

Claude Code, Cursor, Codex. The output ceiling is insane. But "fix the bug" gets you slop. Detailed context gets you shipping code.

The problem: typing out file references, edge cases, and architecture context 40 times a day is exhausting. So you cut corners. Your agent gets lazy input.

Wispr Flow lets you speak full prompts with syntax-aware dictation. camelCase, snake_case, file names all preserved. Auto-tags files in Cursor and Windsurf.

Used by engineers at OpenAI, Vercel, and Clay. Works on Mac, Windows, iPhone, and Android.

Better prompts in. Better code out. Try Free

How to build agents and full-stack apps with Meta Muse: This tutorial walks through Muse Spark, the Muse API, Agent SDK, and Muse Code CLI, then puts them together in real workflows. You’ll learn structured outputs, search grounding, MCP servers, custom skills, persistent memory, autonomous goals, and how to ship a full-stack app with React, Go, and Docker.

Jevgrep (572 ⭐): Jevgrep uses Jev to locate the most relevant files, declarations, and code snippets, then feeds that context directly to Claude Code, Codex, or your other agents so they can start building faster. So you can ask what a codebase does instead of hunting down filenames yourself.

How to build with Claude Sonnet 5.5 (by Anthropic): This guide breaks down when to use Sonnet instead of Opus, how to tune effort for coding and agent tasks, and what to change if you’re migrating from Sonnet 5. It also covers the new thinking behavior, tool-use changes, caching, and why Sonnet 5.5 can cost up to 30% less per task even though token prices stayed the same.

Your CLAUDE.md files, skills, and custom commands were written for older models. It quietly degrades every Opus 5.5 session. AI developer Dan McAteer shared a fix:

It scans your CLAUDE.md, skills, subagents, and slash commands for stale paths, dead commands, contradictions, and prompting patterns written for older models.

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

A new open-source model claims to make agentic coding up to 9× faster without losing performance, and some say it’s better than Jev.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.

We put together a few guides on coding agents, agentic engineering, and leadership frameworks to help you level up in your career. Browse all our guides.

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

Until next time — The Code team