Why Stateful Continuations in AI Agents Make Transport Layers Critical: HTTP vs WebSocket Benchmarks

Got story updates? Submit your updates here. ›

Anirudh Mendiratta examines how the transport layer has become a critical concern for AI agents that engage in multi-turn, tool-heavy workflows. He finds that stateless HTTP APIs suffer from linear payload growth as conversation context accumulates, while stateful WebSocket APIs that cache context server-side can reduce client-sent data by over 80% and improve end-to-end execution time by 15-29%. The performance benefits scale with workflow complexity, making transport layer architecture a first-order concern for the next generation of AI coding assistants.

Why it matters

As AI coding agents become a daily workflow tool for many organizations, the transport layer has emerged as a critical factor in their performance. Stateless HTTP APIs struggle to handle the growing context of agentic coding loops, leading to ballooning payloads and latency. Stateful WebSocket designs that cache conversation history on the server can dramatically reduce overhead, especially for complex workflows involving many tool calls. This architectural innovation has implications beyond a single protocol, highlighting the importance of minimizing redundant data transmission for AI agents at scale.

The details

Anirudh Mendiratta's benchmarks show that WebSocket mode for OpenAI's Responses API can reduce the client-sent data per agentic coding task by 82% compared to the standard HTTP API. This is because WebSocket maintains server-side state, allowing the client to only send incremental updates rather than the full conversation history on each turn. For a typical 10-turn coding task with GPT-5.4, HTTP sent 176KB per task while WebSocket sent just 32KB. This translated to a 29% faster end-to-end execution time. The effect was model-independent, with GPT-4o-mini also seeing a 15% speed improvement via WebSocket. The key innovation is the stateful architecture, not the WebSocket protocol itself - any approach that avoids retransmitting the full context can achieve similar gains.

• In February 2026, OpenAI introduced WebSocket mode for their Responses API.
• OpenAI reports over 1.6 million weekly active users on Codex as of March 2026.

What’s next

As more AI coding agents adopt WebSocket or similar stateful continuation approaches, it will be interesting to see how the performance benefits scale across the industry and impact overall developer productivity.