Developments that matter in production
What changed, how it works and what it means for the people running reliable systems.
40 stories · 13–24
How we reportApigee separates model latency and token use from MCP tool traffic. Tool filters narrow a slow-assistant investigation, while proxy policies determine which charts populate.
Background reading: ObservabilityThe desktop preview packages a configurable coding agent. Its plugins determine which model answers, which files it can read, and where commands run.
Background reading: AI in productionGoogle API Gateway’s streaming preview delivers response chunks while a model is still generating. Creation-time settings and request deadlines shape the migration.
Background reading: Platform engineeringAnthropic’s 50 successful attempts and CAISI’s 61.1% score measure different things. Their task definitions explain what each assessment says about GLM-5.3.
Background reading: AI in productionA persistent assistant can gather context and prepare a dependency patch across conversations. The workspace and installed version determine what its test result means.
Background reading: AI in productionA published Codex Cloud environment gives new tasks prepared files and tools. Existing tasks retain their own working state, including tools installed before an update.
Background reading: Platform engineeringAnthropic reports lower task costs at unchanged token rates. Existing clients must accommodate changed thinking settings, response blocks and forced-tool behavior.
Background reading: AI in productionStandard maintenance has ended for Google’s legacy agents. Migration to Ops Agent changes some metric fields, including disk labels used by existing queries.
Background reading: ObservabilityCLM-8B embeds a situation and ranks supplied actions. Reusable candidate embeddings suit recurring choices, while the candidate list determines which answers are possible.
Background reading: AI in productionStorage Intelligence flags unusual storage usage and previews batch changes. A recurring dataset read shows why the job’s purpose matters when choosing which objects to move.
Background reading: Platform engineeringGoogle’s internal PageBreak agent proposes web vulnerabilities; separate validators test them in running applications. Reproductions can reduce the investigation work handed to developers.
Background reading: AI in productionAlloyDB’s preview gives read-only agent queries separate compute that scales to zero. Shared storage and the first query after idle time shape its value for incident investigation.
Background reading: Platform engineeringDevelopments since our original reporting, with dated updates and supporting sources.
New model documentation makes retained instructions and thinking-block compatibility part of the compaction migration review.
Original report: Claude adds on-demand compaction for long-running agents