Poolside Releases "Laguna M.1" — A 225B MoE Model Specialized for Agentic Coding

Poolside, an AI company focused on coding, has released **Laguna M.1**, an open model optimized for long-running agentic coding, under the **Apache 2.0 license**.

Key specs are as follows:

Architecture**: 70-layer sparse MoE (Mixture of Experts)

Parameters**: 225B total, 23B active

Experts**: 256 (top-k=16)

Context length**: 256K

Designed to interleave reasoning and tool use, and said to excel at long-horizon coding tasks

Poolside also published a 3-bit MLX build for Apple Silicon (M3 Max 128GB), demonstrating roughly 26 tok/s at a peak memory of around 100GB. It's a size that puts local execution within reach even without an expensive GPU cluster.

"Record & Replay" Added to OpenAI Codex — Automate Work Just by Showing It Once

On June 18, OpenAI released a new Codex feature called "**Record & Replay**". A user demonstrates a workflow once on a Mac, and Codex analyzes those actions and stores them as a reusable skill.

It's essentially a "show it once and it's automated" approach that eliminates the need for explicit scripting or prompt tuning. What makes it distinctive is that the AI learns from human work patterns and converts them into reproducible procedures.

Japan's ITmedia also covered it in a breaking report, a sign of how much attention it is drawing at the practical stage.

Agent Dev Tool Competition Intensifies — Cursor, Claude Code, and Devin Pile On

Beyond Codex's new feature, several dev tool companies announced agent automation capabilities on the same day, and competition in the market is heating up rapidly.

**Cursor** launched its "/automate" feature. Describe a task in natural language, and Cursor automatically configures the triggers, instructions, and tools. It also supports Slack emoji triggers, GitHub triggers, and computer use via cloud agents.

**Claude Code** (Anthropic) added an "Artifacts" feature that lets agents output work-in-progress code as shareable live pages. Internally, it is reportedly already being used to share architecture changes and present prototypes.

**Devin** (Cognition) integrated automated security review into its code review feature. It is also rolling out "Devin for Security", which bridges the gap between "finding" and "fixing" in the AppSec domain — security review is fast becoming a first-class task for agents.

This series of moves shows that the center of AI development is clearly shifting from "the model alone" to integrated systems combining "model + harness + memory + SCM".

Artificial Analysis Releases New Benchmark "AA-Briefcase"

Evaluation firm Artificial Analysis has released a new benchmark, "**AA-Briefcase**", modeled on real knowledge work. It is a more realistic design that handles multi-week projects, thousands of fragmented inputs, and corpora of Slack messages, emails, and documents, evaluating tasks such as building financial models and board materials.

Highlights from the results:

| Model | Elo | Cost per Task |

|---|---|---|

| Claude Fable 5 | 1587 | $31.00 |

| Opus 4.8 | 1356 | $10.40 |

| GLM-5.2 | 1266 | $2.40 |

| GPT-5.5 xhigh | — | $3.68 |

Even the top models fully satisfied all evaluation criteria on only **3%** of tasks, underscoring that human intervention remains indispensable for real-world, long-horizon tasks. Meanwhile, GLM-5.2 comes close to the top tier at a fraction of the cost, making it one to watch on cost-performance.

---