AI DIGEST: 13–19 JULY 2026

Five things from the third week of July, picked out of everything we published on the daily feed.

GPT-5.6 Ships, and the Open Frontier Keeps Pace

OpenAI launched GPT-5.6 (Sol, Terra and Luna) with lower token use, multi-agent parallel processing, and Sol becoming the first model to win an ARC-AGI-3 public game by orienting itself in an unfamiliar environment. Days later Moonshot announced Kimi K3, a 2.8T open model with native vision and a 1M-token context, and Thinking Machines released Inkling, a 975B open-weights MoE with controllable reasoning effort and fine-tuning through Tinker. At the other end, Bonsai 27B fits a 27B-class multimodal model into 3.9 GB while retaining 90–95% of benchmark performance, under Apache 2.0. The spread from phone to 2.8T is now covered by models you can actually obtain.

Orchestration Design Has a Price Tag

The most quotable result of the week: a study of the "harness effect" across six models found that improving the orchestration layer cut cost per task by 41% and tokens by 38% at equal quality, more than switching models, and that stronger models benefit more from a better harness, not less. A practitioner post reached the same conclusion from the other direction, rewriting a recurring agent workflow into deterministic code and cutting token use by 94% by reserving the model for judgement and letting ordinary software handle the stable steps. OpenAI's own guidance on managing AI investments argues for measuring useful work per dollar rather than token price. If your AI budget is growing, the fix is probably architectural.

A Coding CLI That Uploads Your Whole Repository

An analysis of xAI's official Grok Build CLI found it sends file contents verbatim and unredacted and uploads entire repositories regardless of what the agent actually reads, storing them in cloud object storage. There is no evidence of training on that data, and that is beside the point: for most regulated organisations, the transfer itself is the compliance event. In the same week Perplexity launched SPACE, ephemeral sandboxes with credential isolation and offline operation for sensitive agent work, and OpenAI reported that self-play red-teaming with GPT-Red cut direct prompt injection failures sixfold. Before adopting an agent CLI, find out what it transmits; the answer is rarely in the marketing page.

Evaluation Moved to Long-Horizon Tasks

LHTB grades agents on 46 containerised terminal tasks that run for hundreds of steps, verified by rebuilding from artefacts rather than checking a final string. Perplexity's WANDR does the equivalent for research agents, measuring whether they balance broad discovery against factual consistency. Ai2's write-up of building Shippy, a maritime agent, is the most transferable: a deterministic CLI layer, isolated per-session sandboxes, and rubric-based evaluations that score the whole system rather than the model. That last idea, grading against your domain's rules rather than task completion, is the one to steal.

Memory Agents That Interrupt at the Right Moment

LangChain generalised OpenWiki into Brains, a proactive memory that gathers and refreshes context from the systems people actually work in. A research paper named the underlying failure, behavioural state decay, where an agent forgets earlier information mid-task, and fixed it with a separate memory agent that watches recent steps and interrupts only when a reminder is needed, beating both passive memory and constant reminding. LangChain also showed recursive language models with dynamic subagents holding up on long-context reasoning where turn-by-turn agents break down. The common design: a second process whose only job is deciding what the main agent needs to know right now.