Inference compute emerges as strategic bottleneck as AI agents reshape coding platforms and harness engineering becomes a key optimization layer

OpenAI is expanding Codex from a coding tool into a general work surface with persistent context, integrations (Supabase, Figma plugin), and team rollout; launched Codex-only seats with $0 seat fee for eligible Business/Enterprise customers through end of June.

WebSocket mode on OpenAI's Responses API keeps state warm across tool calls and yields up to 40% faster agentic workflows; Cursor released an SDK exposing its runtime, harness, and models for use in CI/CD and embedded agents; VS Code shipped semantic indexing, cross-repo search, and prompt evaluation extensions—shifting focus from raw model latency to agent-loop systems engineering and memory retrieval.

Harness engineering research (Agentic Harness Engineering, HALO) shows gains in agent performance through revertible components and trace analysis: Terminal-Bench 2 pass@1 improved from 69.7% to 77.0% in ten iterations, and AppWorld improved from 73.7 to 89.5 on Sonnet 4.6; LangChain's Deep Agents product line introduced Harness Profiles for per-model prompt and tool tuning with built-in profiles for OpenAI, Anthropic, and Google models.