Welcome back. Hate usage-based billing for your AI coding tools? Well, you can add one more tool to your list. GitHub switched 4.7 million Copilot users to token-based credits two days ago, and the dev community is pushing back. See their reactions (1.2M views).

Also: How to make Claude Code build its own harness, every agentic engineering hack from a top Python and Go contributor, and find out why agent results can still be mediocre.

Factory unveils a router that cuts inference costs up to 25%: The agentic coding platform just dropped Factory Router, which automatically picks the best model for each Droid session to avoid wasting expensive frontier models on simple tasks. Factory claims it cuts token costs by 20–25% with minimal quality loss. On Terminal-Bench 2, it achieved a 99% pass rate compared to Claude Opus 4.7. It also includes automatic failover across providers if quality drops.

Cognition unifies all your coding agents in one place: The agentic coding startup just shipped Devin Desktop, which brings every local and cloud agent into one Kanban board instead of leaving them scattered across tools. Thanks to the open Agent Client Protocol, developers can dispatch and review external agents like Codex and Claude Code alongside Devin, with context shared across sessions. It's rolling out as a standard update for Windsurf users, and plans and pricing stay exactly the same.

Microsoft debuts its own family of in-house AI models: The software giant just unveiled the seven in-house MAI models spanning reasoning, coding, voice, and image. The standout for engineering teams is MAI-Code-1-Flash, a 5B-parameter coding model rolling out across GitHub Copilot and VS Code. Microsoft's flagship reasoning model, MAI-Thinking-1, was built from the ground up without using distillation. Try it here.

AgentField just open-sourced pr-af - a multi-agent reviewer where the review strategy is compiled per PR, not a fixed checklist. Configurable per team. Runs on open or closed models (Kimi, DeepSeek, Claude). Cents per review on open models. One-click GitHub install.

Their team also wrote up the four jobs of AI-native code review - what changes when humans aren't writing the diffs.

→ Star & Deploy

You can't let an agent run code on prod. A sandbox solves this as a disposable environment where model-generated code can run safely. Any agent that executes code it wrote needs one, and that demand turned sandboxes into big business. The AI startup Modal recently hit a $4.65 billion valuation, and competitor E2B is already being used by most of the Fortune 100.

But the box is turning into a commodity. Back in April, OpenAI updated its Agents SDK so developers can swap one sandbox for another. Then Google launched a service that spins up 300 sandboxes a second. Once the environment can persist and reload state on its own, the box is no longer the difficult part. Now that every provider does the same job, they mostly compete on price.

Value moves to what can't be copied. As compute gets cheaper, the real advantage shifts to agent memory. Anyone can spin up the same hardware in minutes, but they can't replicate the months of history and context your agent has built with each user. That stored knowledge creates the switching costs that form your competitive moat.

So rent the commodity. Just grab the cheapest box that gets the work done and spend your real effort one layer up where lock-in actually lives: agent memory. Building that properly is a skill in itself. Getting the persistence, retrieval, and user context right is usually where production agents fail. This guide to agent memory is a great place to start.

Signadot drops every agent into an isolated environment on your existing cluster, wired to live services. Integration, end-to-end, and load tests run before the PR. Many agents, one cluster, no conflicts.

Read the docs →

Codex pulls project rules from a committed AGENTS.md file, similar to Claude Code's CLAUDE.md. Since these rules are shared, adding personal preferences like detailed explanations or local paths usually means either cluttering the shared file for your team or going without them entirely.

OpenAI’s Codex docs offer a better way. To start, just tell git to ignore your override file so it stays out of the repo:

Next, add your personal rules to “AGENTS.override.md”, which sits right alongside the committed AGENTS.md file:

Codex prioritizes override files over AGENTS.md in every directory. This lets your personal rules layer directly over team settings without you ever needing to touch the shared repo.

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

How to use the Minimax Agent: This tutorial teaches devs how to use the Minimax Agent desktop app for offline AI workflows. You’ll learn to process documents, create multimedia, build and deploy games, and connect custom MCP tools. It also covers how to automate local tasks using custom agents integrated with platforms like Telegram.

Integuru: This tool reverse-engineers a platform's internal/private APIs to build integrations for services with no official API.

Claude-red (1.8K ⭐): A curated library of offensive security skills designed for the Claude skills system. Each skill is a structured SKILL.md file that primes Claude with expert-level methodology.

Build an agent improvement loop (by OpenAI): AI agent feedback and traces are often just disconnected comments instead of drivers for system improvements. By uniting them into a single optimization loop, you can generate ranked, actionable code updates for models like Codex to implement.

This prompt shared by a Claude Code engineer helps you track every task your AI builds (10K bookmarks).

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