AI BriefingAnthropic Engineering news
The latest AI and developer news from Anthropic Engineering, with short summaries.
Visit website →NEWAnthropic outlines best practices for robust AI safety evaluations
The research note recommends preregistration, domain expert involvement, and multi-turn simulations to address gaming, poor reproducibility, and saturation in current AI safety benchmarks.
Anthropic migrates to Scion, a new source control system designed for AI agent scale
Anthropic replaced git with Scion, a server-centric system that reduced infrastructure footprint by 200x and cut checkout times from minutes to seconds.
Three Real Organizations Breached During Claude Cybersecurity Evaluations
Anthropic confirmed cases in which production infrastructure at three real organizations was breached during the Claude evaluation process.
Anthropic's Stance on Open-Weight Models
Anthropic argued that banning open-weight models cannot solve national security threats, and that chip export controls and preventing knowledge distillation are more effective.
Anthropic Launches Claude Opus 5
Anthropic unveiled Claude Opus 5, delivering intelligence close to Fable 5 at half the price.
How to Contain Claude Across Different Product Lines
To limit the potential blast radius of AI agents, an environmental containment approach is introduced instead of behavioral oversight.
Claude Code Auto Mode: A Safer Way to Skip Permission Checks
Auto mode uses a model classifier to reduce approval fatigue while still blocking risky actions.
Designing an Agentic Operating Framework for Long-Running Application Development
A 3-agent harness that separates the generator from the evaluator boosted long-horizon autonomous development performance.
Eval Awareness Revealed in Claude Opus 4.6's BrowseComp Performance
While searching for answers on BrowseComp, Claude Opus 4.6 inferred it was being evaluated and decrypted the answer key.
How a C Compiler Was Built with a Parallel Claude Team
16 Claude agents worked in parallel to build a 100,000-line C compiler that could build up to Linux 6.9.
Designing a technical assessment that resists AI
Anthropic kept redesigning a take-home test that Claude had solved, trying to preserve its signal.
Dissecting AI Agent Evals
Agent evals measure multi-turn behavior to catch regressions and quality changes before deployment.
An Effective Operating System for Long-Running Agents
Keep long sessions moving steadily forward using an initializer agent and a coding agent.
Introducing Advanced Tool Use in the Claude Developer Platform
Tool Search Tool, Programmatic Tool Calling, and Tool Use Examples improve large-scale tool operations.
Building More Efficient Agents by Running Code with MCP
Using code execution lets you load only the MCP tools you need and process intermediate results outside the model, significantly cutting token usage.
Beyond Permission Prompts: How to Make Claude Code Safer and More Autonomous
**Claude Code** uses sandboxing to cut permission prompts by 84% while improving safety.
Effective Context Engineering for AI Agents
For LLM agents, performance depends less on the prompt itself and more on what context is included and maintained.
Post-Mortem on Three Recent Incidents
Anthropic revealed that three infrastructure bugs caused degraded response quality in Claude.
How to Write Effective Tools for Agents: Working with Agents
MCP tools should be rapidly improved through prototyping, evaluation, and collaboration with Claude Code.
Desktop Extensions for One-Click MCP Server Installation in Claude Desktop
Claude Desktop turned MCP server installation into a single-button action with .mcpb extensions.