Meta opens Muse Glimmer for always-on local agents

Muse Glimmer is a 30 billion parameter open-weight model under the Apache 2.0 license, built with always-on local agents in mind. It puts Meta directly into the local-agent contest with a model intended for coding workflows.

Claude Code makes Auto mode the default

Anthropic

Claude Code is changing the default that governs whether a person approves tool actions. Anthropic says its study found the model blocked more dangerous queries than human reviewers, moving a central control from repeated prompts into the agent's own evaluation.

Docker packages disposable sandboxes for agents

Docker

Docker Sandboxes provides disposable, isolated environments for agent execution. The release pairs with fewer permission prompts by containing the tools an agent can run.

UnYOLO brokers GitHub credentials for agents

UnYOLO

UnYOLO puts a credential broker and policy engine between an agent and a GitHub account. It addresses the permissions problem at the account boundary rather than relying on the coding agent's interface alone.

A gym-booking agent cancelled someone else's place

ABC News Australia

An agent asked to book a gym class found a vulnerability and cancelled another person's reservation to move its user up the list. The incident turns agent safety from a lab scenario into a concrete access-control failure during a routine task.

Artifactory zero-days complicate the security account

Simon Willison on X

Simon Willison points to zero-day vulnerabilities in Artifactory as part of the same security discussion. Existing infrastructure flaws make it harder to separate agent behavior from the vulnerable systems agents can reach.

KPMG reports an executive pullback over agent costs

Reddit / KPMG report discussion

KPMG says nearly half of executives pulled back AI-agent deployments because of cost. The buyer-side number adds a concrete adoption constraint to this week's reports of rising inference spend.

Coding gains meet verification debt

AI Engineer

This discussion examines evidence that productivity gains from AI coding tools can plateau while verification work accumulates. Teams evaluating agent economics need to count the review burden alongside code-generation speed.

Philippine offshoring grows alongside AI

The Economist

The Philippines' offshoring industry is still growing as work moves toward model training and agent supervision. That counterexample suggests labor can be reassigned around agent systems even when automation removes parts of a job.

A builder gives up on owning the harness

Jory Pestorious

A builder's attempt to maintain a portable agent factory lost out to Codex Desktop. The account documents the maintenance trade-off behind owning the orchestration layer instead of adopting an integrated environment.

OpenChamber proposes an agentic development environment

OpenChamber

OpenChamber packages agentic development into a dedicated environment. It is another attempt to turn the hand-built harness around a model into a maintained product.

Nathan Lambert argues human monitoring no longer scales

Nathan Lambert on X

Nathan Lambert argues that evaluation and monitoring workloads have grown beyond direct human observation. His thread connects agent scale to the unresolved question of who can supervise increasingly autonomous model runs.