ARCHITECTURE

OpenAI published the engineering story behind GPT-Live, its third-generation voice system: a full-duplex model with no turn detector, Go replacing Python on the media path, seamless stateful handoffs, and WARP, a new WebRTC transport going through the IETF.

MCP gives an agent live access to tools and data. Agent Skills give it packaged procedure. They solve different halves of the same problem, and the MCP working group is now standardizing how skills ship over MCP. Here is the decision rule.

A new paper proposes inverting traditional agent architecture - making the append-only event log the source of truth, not an afterthought. HN debates whether this is novel or just CQRS with extra steps.

Dan Abramov's explainer on ATProto architecture is making the rounds. The core insight: Bluesky's protocol separates hosting from applications in a way that Mastodon-style federation fundamentally cannot. Here's what that means for developers.

A practical architecture for multi-step Claude agents. Loop patterns, state management, error recovery, and the production gotchas that turn a five-step demo into a 20 percent success rate at scale.

The math of agent pipelines is brutal. 85% reliability per step compounds to about 20% at 10 steps. Here is why long chains collapse in production, and the six patterns the field has converged on to fight the decay.

Martin Fowler reframes AI-era debt into three layers - technical, cognitive, and intent. The third one is the one most teams are silently accumulating. Here is what it is and how to diagnose it.

From single-agent baselines to multi-level hierarchies, these are the seven patterns for wiring AI agents together in production. Each with a decision rule, an implementation sketch, and the tradeoffs that actually matter.

CloudFlare, Anthropic, and Cursor independently discovered the same pattern: don't load all tools upfront. Let agents discover what they need. The results are dramatic.