The Agent Chronicles

The day in AI, in brief.

The AgentChronicles
Earth, on

Agents, models and API security

Browser-operating agents and faster open models arrive as research raises fresh questions about reasoning-trace exposure and reliable enterprise control.

Briefs

Models and platforms

  • NVIDIA releases Nemotron 3.5 Lightning — NVIDIA has released Nemotron 3.5 Lightning, an open mixture-of-experts model with 30 billion parameters and three billion active. NVIDIA claims up to four times faster output for specialised, always-on agent workloads.

  • Hetzner trials free open-model inference — Hetzner is offering a free experimental open-weight inference API. Developers can test workloads and provide feedback while the service remains exploratory.

  • Gemini passes one billion monthly users — Google says the Gemini app now exceeds one billion monthly users. The company calls it its fastest-growing product and its 14th to reach that threshold.

  • OpenAI previews ChatGPT and Codex for Linux — OpenAI is previewing a Linux desktop app that brings ChatGPT, ChatGPT Work and Codex to supported systems. It gives Linux-based organisations a native route into OpenAI’s chat and coding workflows.

  • Quantised Qwen inference matches across hardware — A hackathon demonstration produced matching int8 output hashes for Qwen3-0.6B across A100 and H100 GPUs, plus several CPUs. Reproducible runs could simplify testing and auditing across mixed hardware.

Agent engineering

  • One control plane for analytical agents — The proposed Control Plane Pattern gives one central agent responsibility for judgement. Conventional code detects anomalies, while a knowledge graph limits the hypotheses and data joins the agent can pursue.

  • AI coding failures move up the stack — Developer commentary argues that AI coding failures now cluster around architecture, usability and missing product context. Teams should put more effort into design review and adversarial testing.

  • Write commit conventions for coding agents — The discussion on AI-written pull requests and clear commit-message guidance give coding agents useful limits: concise subjects, imperative verbs, and an explanation of what changed and why. Shared conventions also make reviews and later debugging easier.

  • Omarchy enables unattended coding agents — Omarchy has merged permission-bypass launches for nine coding agents running unattended. The change speeds desktop automation, but removes a human approval point before agent actions.

Security and business

  • Research probes reasoning-trace leakage in model APIs — Researchers claim proprietary model APIs can expose hidden reasoning, with recovered token counts often matching billed thinking tokens. A separate tool-based demonstration shows how visible reasoning may be elicited despite disabled internal thinking.

  • Decagon’s productisation test for deployment teams — Decagon argues in its essay on forward-deployed engineering that deployment teams should discover new workflows, then turn recurring configuration and tuning into product features. It says Duet now handles two-thirds of deployment work autonomously.

  • River AI announces $1.1bn funding round — River AI says it has raised a $1.1 billion round led by General Catalyst and AMP. NVIDIA, AMD, Y Combinator and Temasek are among the named backers.

Automation in practice

  • Grok Bot begins browser-agent beta — Grok Bot is entering beta as a set of agents that sign into business tools and complete work through browser interfaces. Its design makes permissions and human oversight central to safe use.

Interviews & Talks

Why agent teams may need a flexible harness

A founder argues in this interview and demonstration that stronger models have made rigid agent graphs less useful. The proposed alternative is a general-purpose harness: a virtual machine with memory, evaluations, command-line access and links to company tools. In this view, agents should take on broader goals rather than pass work through fixed lattices of specialised workers. That can make systems easier to adapt as models improve. It also puts more weight on the surrounding harness. Teams need clear tool permissions, useful evaluations and ways to contain failure when an agent can act across internal systems. The discussion is a founder’s account, but it captures a live design choice for teams building agent systems.

Worth watching

  • 00:31 — The speaker explains why newer models made the earlier graph-and-lattice design feel constraining.
  • 00:51 — The founder describes replacing two years of working infrastructure with a flexible harness.
  • 01:14 — The proposed harness combines a virtual machine, memory, evaluations and command-line access to company tools.