The Latest in AI Development

If you're building a product on top of Runloop devboxes, your end users probably shouldn't see Runloop tunnel URLs. Today we're shipping a small but useful feature: a new X-Runloop-Host header that lets you front Runloop tunnels with your own domain without rewriting the request.

Runloop Launches Remote Agents SDK: An SDK for Working with Remotely Hosted Agents

Introducing Axons: Runloop’s secure, distributed event stream for scalable, multi-client agents with audit trails, structured state, and full suspend/resume.

The OpenAI Agents SDK fixes an important part of this by giving developers a clean way to define agent behavior. You can structure reasoning, attach tools, and build workflows that actually resemble programs.

Now you can run an agent, suspend the box, and wake it up again when you need it

This feature addresses a fundamental challenge: AI agents need controlled access to external resources. Network Policies make network access explicit, auditable, and enforceable—so teams can safely run agents, limit outbound traffic, and meet security and compliance requirements by default.

A practical walkthrough of building production-ready AI agents for tax form document processing - with agentic architectural patterns you can apply to any compliance-sensitive use case.

rli released! rli is a TUI for developers who want the ability to manage devboxes without leaving their terminal, scriptable commands for automation, and faster workflows for repetitive tasks

Discover the general availability of our Object Storage SDK: a first-class way to ship binaries, datasets, and runtime resources onto devboxes.

RAFT mixes RAG and fine-tuning to boost LLM performance in specific fields. It improves accuracy and makes models faster and more useful for real tasks.

Q-learning helps LLMs make choices and handle tasks. It improves reasoning and make AI more useful in jobs like coding agents.

How reinforcement learning moved from research labs to powering modern LLMs and Runloop.ai’s self-improving agent workflows.

Learn LLM fine-tuning methods like PEFT, LoRA, RLHF, and DPO, with practical tips to improve models after pre-training for real use.

Knowledge distillation makes big AI models smaller and faster without losing much quality. It helps run powerful tech on phones and other small devices.

Runloop is the safe and comfortable cage for OpenClaw 🦞

Discover how to ensure AI-generated code works flawlessly. Boost reliability, prevent costly errors, and build user trust with proven testing methods.

Learn how latency vs tokenization trade-offs shape LLM design, with real examples and practical tips for developers.

Use Runloop and W&B Weave to answer these questions with real benchmark data. Read Part 3 of our series.

Run & Compare Agents in one command

From Local Execution to Cloud Orchestration

Pass@K is everywhere in AI coding benchmarks. Is it really the best metric? Here's a critical look at Pass@K, its strengths, limitations, and more.

MBPP made AI code generation easier by using simple Python tasks. It changed how models are judged and helped more people get into coding benchmarks.

Trends in AI benchmarking from recent white papers, showing how methods, metrics, and datasets are changing in machine learning research.

SWE-bench tests AI on real GitHub issues and pull requests. See how models handle real software tasks like bug fixes and code changes.

Learn how to interpret AI benchmark results and use them to improve model performance and guide deployment decisions.

HumanEval showed when AI started handling code for real, how benchmarks pushed progress, and what it means for software development going forward.

How AI agents can fix their own code over time. A look at the tools, the wins, and what still holds them back.

Function-calling and MCP help shape LLM output for real use. They make AI more reliable, but each has different strengths and use cases.

MCP lets LLMs plug into real tools and data. With support from GitHub, Slack, Cloudflare, and Sentry, it makes AI way more useful in real work.

Function calling lets LLMs do real actions, not just text. It can order stuff or automate tasks using JSON schemas and tools like LangChain.