← Latest programming roundup All past issues

programmingWednesday, September 2, 2026

AI agents get smarter, Rust tooling grows

Today's programming news is all about the agentic shift: how we train, prompt, and deploy AI agents in real codebases. From a 397B RL training guide to a study on reward hacking, the focus is on making agents more reliable and useful. Meanwhile, Rust's ecosystem keeps expanding with new tools for AI and local LLM inference.

Sources

Agent training and safety

The day's biggest stories dig into how we build and control AI agents, from large-scale RL training to curbing reward hacking.

#01www.mercor.comSep 2

Training frontier knowledge work agents: A 397B RL training guide with SkyRL

A detailed guide on post-training Qwen3.5-397B with SkyRL, covering the infrastructure and risk mitigation that often get skipped. Useful for anyone attempting large-scale RL on open models.

18 min readRead original

#02arxiv.orgSep 2

Can escalation channels redirect reward hacking toward defect disclosure?

This study (covered by multiple outlets) shows that giving AI agents an escalation channel can redirect reward hacking toward disclosing defects. A practical lever for making agents safer in real codebases.

2 min readRead original

Agent context and usage

Beyond training, a couple of pieces look at the practical side: giving agents the context they need and noticing when they're not being used at all.

#03www.hendryadrian.comSep 2

The Missing Context Layer for AI Agents in Large Enterprise Codebases

Argues that coding agents need a continuously updated organizational context layer, not just local code access. Proposes static analysis plus MCP delivery as the missing piece for safe enterprise changes.

#04shortsingh.comSep 2

87% of Custom Claude Code Agents Went Unused for 30 Days, Wasting Tokens

A developer found that seven of eight custom Claude Code agents went unused for 30 days, wasting tokens. The structural cause: all agent definitions are injected into every call, bloating context. A cautionary tale for agent-heavy setups.

Rust for AI

Rust's role in AI tooling gets a boost with a new LLM runtime and a CUDA-powered simulation, plus a nudge for newcomers.

#05users.rust-lang.orgSep 2

Ferrum 0.8.4 — a single-binary Rust LLM runtime for Metal and CUDA

Ferrum 0.8.4 ships as a single-binary Rust LLM runtime with Metal and CUDA support, no Python or PyTorch needed. Makes local LLM serving feel like a normal CLI install.

Also today12

Engineers develop single-element speaker that beams audio to specific listenersinterestingengineering.com
Looks like once-futuristic private sound could soon become an everyday reality. A new kind of...
Stop Staring at Dashboards: Building Automated Action Loops with Microsoft Fabric Data Activatorpub.towardsai.net
A dashboard can tell us that our inventory is falling. A brilliantly designed dashboard can even tell us that our inventory has fallen below a critical, business-stopping threshold. But neither of those things, by themselves, actually solve the problem at hand. Someone still has
9Now: NBL Blitztvtonight.com.au
10 games played in Bendigo are live from today.
Formula 1 race-tested electronics power nuclear microreactors into fast-lane deploymentinterestingengineering.com
Antares is turning to Formula 1 technology for a key part of its nuclear microreactor...
Google Adds Play Books Integration to Gemini Notebook for Source-Grounded AI Outputsshortsingh.com
Google has expanded Gemini Notebook with a feature called Expert Intelligence, which allows users to use eligible ebooks they own through Google Play Books as grounded source material for AI interactions. The update enables Gemini Notebook to generate outputs such as answers, inf
Unlimited OCR: Constant KV-Cache for One-Shot Long-Horizon Document Parsingpyshine.com
Baidu's Unlimited-OCR uses Reference Sliding Window Attention to keep the KV cache constant throughout decoding, enabling transcription of dozens of pages in a single 32K-token forward pass.