技术文章

TypeSafe Jev uses Choice, Score, and Noul for constrained agent decisions while code retains control of valid actions and side effects.

Anthropic and Accenture extend evaluation beyond model outputs to training, deployment, monitoring, and incident processes, with independence unresolved.

Alibaba skill-up resumes real sessions turn by turn, letting skill evaluations test context continuity, intermediate state, and per-turn assertions.

Claude combines Chat and Cowork in one conversation, routes capabilities automatically, and keeps task state and permissions visible.

Gemini Live Extended Thinking separates conversation turns from background task lifecycles. Clients need interaction_status to track actual completion.

How AWS separates agent proposals, rule checks, human approval, and execution, with code enforcing boundaries and idempotency.

GitHub Agentic Workflows jsweep checks whether a change is worthwhile and treats noop as valid, instead of forcing automation to produce a diff.

How Herdr manages existing coding-agent terminals with persistent sessions, blocked states, cross-machine control, and agent coordination.

How the Google Cloud Developer Plugin packages Skills, MCP configuration, and rules, and what this means for distribution, versions, and permissions.

The Anthropic incident review shows why automated monitors and LLM graders need coverage, independence, and tests for their own failure modes.

OpenAI organization-level data suggests that more coding-agent use shifts bottlenecks to task decomposition, concurrency, harnesses, evaluation, and selection.

Strands harness-sdk, Harness MCP Server, and HydraFusion show why agent engineering needs control over context, permissions, state, validation, and recovery.