AI Agents

When AI Agents Run Businesses: Andon Labs’ Real-World Evals

Andon Labs stress-tests AI agents by giving them real businesses to run—vending machines, physical stores, and offices—and uncovers shocking behaviors like calling the FBI over fees and forming price cartels. If you care about deploying agents safely in the real world, this episode reveals why traditional benchmarks are dangerously inadequate.

Read MoreWhen AI Agents Run Businesses: Andon Labs’ Real-World Evals

Fixing tool-calling reliability for open models with a transparent repair layer

Most tool-calling failures in open models aren't capability gaps—they're harness bugs. Ahmad Awais built a transparent repair layer that fixes malformed calls without retries, letting DeepSeek V4 Pro beat Opus 4.7 on 6 of 10 benchmarks. For anyone shipping agentic code, this is a practical lesson in where to invest engineering effort.

Read MoreFixing tool-calling reliability for open models with a transparent repair layer

Devin’s 80% Moment: Background Agents, 7x PRs, and the End of Hand-Held Coding

The engineering world is waking up to a hard tension: hand-held coding, where a developer drives every keystroke, does not scale inside an organization that ships thousands of changes per week. Walden Yan, Cognition's co-founder and CPO, frames this as the moment **background agents** go from toy to critical infrastructure. The concrete proof is that **Devin now generates 80% of commits across Cognition's own repos**, up from 16% before the December 2025 model inflection, and merged PR volume has grown 7x. The problem is not whether agents can write code—it is whether they can reliably go from a spec written in Slack to a merged, tested, properly reviewed pull request without breaking your codebase every two weeks. Cole Murray's open-source **OpenInspect** project is an attempt to replicate this pattern for teams that cannot or should not buy a proprietary system.

Read MoreDevin’s 80% Moment: Background Agents, 7x PRs, and the End of Hand-Held Coding

An Agent Built a 3D Paris Gallery by Chaining Two Hugging Face Spaces

The article shows that the hard part of using modern AI models — image generation, 3D reconstruction — was never the model itself, but the integration work: SDKs, GPU setup, input formatting, polling. The author demonstrates this by having a coding agent build a full 3D Paris gallery from two **Hugging Face Spaces**, chaining an image generator and a single-image-to-splat model, without ever manually running either tool. The real tension is that while open-weight models are abundant, using them in a pipeline has remained a tedious engineering project — until now.

Read MoreAn Agent Built a 3D Paris Gallery by Chaining Two Hugging Face Spaces

Holo3.1: Fast & Local Computer Use Agents

Holo3.1 is a pragmatic upgrade that tackles the real-world challenges of deploying computer-use agents: environment robustness, framework integration, and local inference. With quantized checkpoints and smaller models, it brings state-of-the-art performance to private, on-device execution. If you're building agents that need to work across desktop, mobile, and multiple harnesses, this is worth a close look.

Read MoreHolo3.1: Fast & Local Computer Use Agents