AI Briefing#agent-architecture
The latest AI and developer news about #agent-architecture, with the original source and a short summary.
The latest AI and developer news about #agent-architecture, with the original source and a short summary.
The architecture relies on event inboxes, timers, memory, permissions, and outboxes to manage state across unattended runs.
The agent generated 1,540 artifacts in 12 weeks, saving 337 hours of rep time by using business context rather than just prompt tuning.
The feature is available in Python and JavaScript, with task notifications currently in beta.
The new browser tools, built on Stagehand, outperformed six other toolsets in benchmarks, with the closest competitor costing 9x more.
Anthropic advises defaulting to one model and context window, noting that tool choice degrades past 30-50 definitions and that a single loop ran 2x faster than multi-agent harnesses.
The study benchmarks KV cache limits and conversation state management strategies for agent model switching.
Anthropic's Applied AI team advises on when to fold subagents into the main agent loop, noting that single-agent designs with skills frequently outperform multi-agent architectures in quality and cost for specific enterprise deployments.
The pipeline generated over 1 million lines of code in five months for $135,000 in Claude token costs, achieving a 164x speedup for Db2 migration compared to previous efforts.
The guide recommends a single-agent architecture with skills over subagents to reduce latency and cost, supported by a GitHub reference implementation.
LangChain adds an interpreter to Deep Agents, providing a middle layer between tool calls and sandboxes
Proposes an architecture that separates the execution loop from the sandbox to improve the security and scalability of AI agents.
It's an API pattern where Sonnet/Haiku executes and Opus only advises when the process gets stuck.