Tech • IA • Crypto
Anthropic Claude Co-Work, Memory & Opus 4.7 Redefine Agents
Briefing
Claude Co-Work turns AI operator
Anthropic introduced Claude Co-Work, shifting AI from assistant to autonomous operator. The system executes workflows across local files, Google Drive, Notion, Slack, and Chrome with user approval. It generates structured plans before acting, outlining steps and expected changes. This marks a major move toward delegated, end-to-end task execution.
Opus 4.7 scales test-time compute
Claude Opus 4.7 demonstrates major gains through test-time compute scaling, increasing reasoning effort during inference. Higher effort runs used up to 10× more compute, producing significantly better outputs in complex simulations. A baseline run took 50 seconds and 4,600 tokens, while extended runs doubled resources for improved realism. The trade-off shifts optimization toward balancing latency, cost, and performance.
Anthropic unveils persistent AI memory
Anthropic introduced a file-based memory system enabling agents to continuously learn from experience. Agents store success patterns, failures, and strategies in a structured hierarchy they can update over time. Combined with a process called Dreaming, systems refine behavior without retraining. This enables long-lived agents that improve autonomously across tasks.
Google Cloud demos Claude agents
Google Cloud showcased end-to-end development powered by Claude-based agents. A single system handled ideation, design, coding, deployment, and analytics using tools like Cloud Run, Firestore, and BigQuery. A structured “plan mode” ensures architectural alignment before execution. The demo highlights collapsing traditional role boundaries across software teams.
Asana builds multi-agent enterprise workflows
Asana is deploying collaborative AI agents with shared organizational memory and context. These agents operate within a structured work graph linking goals, tasks, and roles. Persistent memory allows agents to retain institutional knowledge, even after human creators leave. The approach enables coordinated multi-agent execution across complex business processes.
MCP becomes agent integration standard
Model Context Protocol (MCP) is emerging as a standard for connecting AI to tools and data sources. It supports integrations across services like Linear, documentation systems, and productivity platforms. Two architectures—HTTP servers and STDIO servers—enable both remote and local execution. MCP underpins the shift toward tool-using, action-oriented AI systems.
Replit ByBench tracks real-world coding
Replit launched ByBench, an open benchmark for evaluating AI agents building apps from scratch. It uses around 20 real-world PRDs and automated agents to test functional correctness. The system replaces static benchmarks with continuous evaluation using production data. This reflects growing demand to measure real-world performance of “vibe coding” systems.
AI shifts engineering bottlenecks downstream
AI-assisted coding is moving bottlenecks from writing code to verification, security, and coordination. Teams are replacing long-term planning with just-in-time decision-making and rapid iteration. Working prototypes increasingly resolve debates instead of design discussions. This forces organizations to rethink workflows built around scarce developer output.
Videos covered
Getting Started with Claude Cowork
- •AI Moves From Assistant to Operator
- •Deep Integration Across Environments
- •Controlled Access and Permissions
Running an AI-native engineering org
- •Bottlenecks move beyond coding
- •Traditional processes “quietly stop working”
- •Planning shifts to just-in-time models
The thinking lever
- •Test-Time Compute Emerges as Key Scaling Method
- •Performance Gains Demonstrated with Effort Scaling
- •Three Types of Tokens Drive AI Behavior
Building with Claude on Google Cloud
- •End-to-end AI-assisted development
- •Rapid prototyping from simple inputs
- •Structured planning before coding
Evaluating and improving Replit Agent at scale
- •Shift from static to continuous evaluation
- •Unique demands of “vibe coding”
- •Introduction of ByBench
Giving coding agents their own computers: How Cursor built cloud agents
- •Bottleneck shifts from models to infrastructure
- •Three-stage evolution of agent systems
- •Onboarding agents like human developers
Memory and dreaming for self-learning agents
- •Memory as a Core AI Primitive
- •From Static Context to Self-Learning Agents
- •File-System-Based Memory Design
Building with Claude Managed Agents and Asana AI teammates
- •Toward the “agentic” enterprise
- •Current limits of enterprise agents
- •A shared enterprise memory
The expanding toolkit
- •Shift From Models to “Toolkits”
- •End of Manual Tool Routing
- •Smarter Tool Design Improves Performance
MCP in Claude Code
- •What MCP Does
- •Shift Toward Tool-Driven AI
- •Wide Range of Integrations