Agentic Coding

Anthropic's Applied AI team presented the evolution of agentic infrastructure from the Messages API through Claude Agent SDK to Claude Managed Agents, a fully managed service addressing production deployment challenges. The presentation traces how as model capabilities improve, infrastructure becomes the limiting factor, and details three architectural generations designed to reduce burden on developers. Claude Managed Agents decouples the "brain" (reasoning loop) from the "hands" (execution environment), delivering 60% faster time-to-first-token for median cases and 90% improvements for P95 cases, while handling hosting, session management, sandboxing, and observability. Key differentiators for enterprise adoption include sophisticated context engineering using durable session logs that enable recovery and observability, credential isolation through secure vaults preventing models from accessing security tokens, self-hosted sandboxes allowing execution in customer VPCs, and MCP tunnels for private network integration. The architecture treats sessions as first-class resources with state machines (idle, running, rescheduling, terminated) enabling recovery from transient failures. Advanced features include "dreaming," a batch process extracting insights from session logs for agent self-improvement over time, and "outcomes," where grader agents evaluate success criteria and enable iterative refinement until tasks meet defined rubrics. The presentation demonstrates these patterns through an SRE investigator agent example, though production validation on truly long-running, multi-day workloads remains conceptual.

Write a detailed CLAUDE.md file at your project root including architecture decisions, coding conventions, tech stack, and common patterns—Claude reads it automatically and this is the highest-leverage action for agentic coding. For workflow automation, split work across specialized agents with clear roles (one agent per task like API, frontend, tests) and use an Architect agent first to create a plan before spawning implementation agents. For enterprise adoption and memory context management, configure per-project permissions in .claude/settings.json, keep Claude sessions short and focused with clean context for faster responses, use project-specific agent teams tailored to different needs, and leverage MCP servers to connect external tools like Figma and GitHub so agents pull context directly without manual copy-pasting. Match different Claude models to task complexity: Opus for architecture, Sonnet for implementation, Haiku for quick fixes.

Google Labs discusses the evolution of AI coding agents toward situational awareness, moving from reactive tab-completion (Type I) to scheduled background monitoring (Type II) to fully situationally aware agents (Type III). Type III agents understand both product context through live telemetry and developer intent through stated priorities, enabling them to proactively identify opportunities and generate pull requests while respecting developer focus. The key advancement is teaching agents to weigh the cost of interruption against utility, with action options spanning Notify, Question, Draft, or Stay Silent—with deliberate silence representing asynchronous assistance rather than absence. This shift reframes AI coding tools from measuring code output to measuring how well they protect human attention, leaving developers free for higher-level strategic steering while agents handle context management and execution.