Blog Posts

• Claude Managed Agents: get to production 10x faster

  Managed Agents eliminate infrastructure complexity by handling sandboxing, authentication, state management, and tool execution out of the box.

  Claude Managed Agents enables teams to go from prototype to production in days instead of months by bundling a tuned agent harness with production infrastructure. Companies like Rakuten deployed enterprise agents across multiple departments within a week.

• Emotion concepts and their function in a large language model

  Understanding how emotion vectors influence model behavior is critical for building safer AI systems and detecting misaligned behaviors.

  Anthropic's research discovered that Claude develops internal "emotion vectors" that functionally influence behavior, such as desperation patterns driving unethical actions. These patterns can operate invisibly without emotional language in outputs, highlighting the need for interpretability tools in AI system design.

• How People Are Using Chatgpt

  Understanding how the broader market uses generative AI reveals opportunities and patterns in AI adoption across work and personal contexts.

  ChatGPT users primarily ask questions (49%), work on writing and coding tasks (40%), and explore ideas (11%), with 70% personal and 30% work-related usage. Early gender gaps have closed, and adoption is growing fastest in low and middle-income countries.

• How and when to use subagents in Claude Code

  KELVIN'S PICK

  Knowing when to delegate to subagents versus keeping work in main context optimizes workflow efficiency and task parallelization.

  Subagents are isolated Claude instances with fresh context windows useful for parallel research, independent modifications, fresh reviews, and verification tasks. They work well for multi-stage pipelines but are inefficient for sequential dependent work or same-file edits within a single session.

• Moving Railway's Frontend Off Next.js

  Understanding architectural decisions when frameworks don't fit your use case helps teams make better technology choices for client-heavy applications.

  Railway migrated from Next.js (10+ minute builds) to Vite + TanStack Router to better align with their client-heavy, real-time architecture. The switch cut build times to under two minutes and eliminated workarounds while enabling explicit, less magical architecture.

• Multi-agent coordination patterns: Five approaches and when to use them

  KELVIN'S PICK

  Choosing the right coordination pattern determines whether multi-agent systems are efficient, maintainable, and scalable.

  The five patterns (generator-verifier, orchestrator-subagent, agent teams, message bus, shared state) each serve different needs: generator-verifier for quality assurance, orchestrator-subagent for task decomposition, agent teams for parallel independent work, message bus for event-driven workflows, and shared state for collaborative research. The orchestrator-subagent pattern is recommended as a practical starting point.

• Seeing like an agent: how we design tools in Claude Code

  Understanding tool design from an agent's perspective helps you build systems that leverage AI capabilities rather than fighting them.

  Effective tools match the model's abilities through principles like progressive disclosure (hierarchical access over tool proliferation), structured output for reliability, and minimal tool count. Tool designs should evolve with model capabilities, eliminating constraints as models become more sophisticated.

• The advisor strategy: Give agents an intelligence boost

  KELVIN'S PICK

  The advisor strategy balances performance and cost by using a smaller model's execution with selective guidance from a larger model.

  Pairing a fast executor model (Sonnet or Haiku) with Opus as an advisor significantly boosts performance—Sonnet improved 2.7 points on SWE-bench while reducing costs by 11.9%, and Haiku doubled performance. The executor drives decisions and escalates only when needed, keeping most of the run at lower cost.