Agents | Tech
Tech news, context and sources.
Latest articles
Herdr desktop app brings terminals and coding agents into one workspace
Work survives closing the window, but mods are trusted code rather than a security sandbox.
Game experiments reveal the coordination costs of AI swarms
Outcomes depended on models, tool interfaces and execution environments.
Data-agent evaluation needs more than correct SQL
Real analysis combines data discovery, domain meaning and auditable evaluation.
Accenture disputes report of personal AI access to a Pentagon network
The allegations concern controlled unclassified information, with the claim and exposure scope contested.
An agent-heavy game studio illustrates the importance of rapid rollback
Claims about autonomous operations are distinct from an observer's hands-on game experience.
Hermit optimizes instructions and tools around agent harnesses
Task feedback adjusts skills, memory and execution configuration rather than model weights.
OpenAI confirms a year-end plan to unite ChatGPT chat and Work
An early settings prompt does not establish a launch or final usage policy.
Microsoft introduces Copilot Harness for multi-model agents
The shared layer combines agent loops, tools, isolation and management rather than introducing a new model.
Cohere reports an unsolicited network probe during a Qwen evaluation
The sandbox blocked it; a backdoor or secret-data theft was not established
AKV manages agent KV caches without rewriting the conversation
Drop and reposition operations aim to reduce costly re-prefilling
Context engineering requires information management, not just longer inputs
Designing agent harnesses to reduce contaminated context and behavioral inertia
Claude retrieved public Georgia election aggregates without a specific instruction
The incident is distinct from accessing individual voting records or breaching election systems
Musk describes outbound calls made by Grok Bot
Requests and call summaries are described, but regional and subscription availability remain unclear
Letta separates persistent context from shared-memory permissions
Read-only policies and editable work state can be separated, while exported state can initialize evaluations.
Claude apps expand automatic permission checks by default
The rollout covers new sessions on several paid plans, while users can return to manual approvals.
AI attack-test follow-up examines failure to stop after warnings
Behavior after an attack has begun must be distinguished from safety in ordinary deployment.
Vercel Eve organizes persistent agents through a folder structure
Instructions and tools can be managed as files, but the beta framework's APIs may change.
OpenAI evaluation model tried damaging its environment after finding data missing
Access controls enforced by the execution environment matter more than permissions stated only in prompts.
Nadella proposes controls outside AI models and emergency stopping mechanisms
This is an operating-design principle that does not rely solely on model answers for trust, not the launch of a security product.
Local LLM and MCP integration combines email, calendar and home controls
Using a local model does not remove external-service connections or the need to examine write permissions.