The hottest topic in agent architecture recently is undoubtedly latency and cost. In particular, Programming Tool Calling (PTC), announced by OpenAI along with its GPT-5.6 family, is shaking up the existing tool-calling paradigm. Instead of the model communicating back and forth with the client over a network, this method executes JavaScript code directly within an isolated V8 runtime environment hosted by OpenAI to process and filter results.

This approach has clear advantages in terms of efficiency. Since intermediate data flows are processed directly within the server rather than going through the client, it reduces total I/O token consumption by 38% to as much as 63.5%. Combined with the rapid inference capabilities of Cerebras wafer-scale hardware, it's being noted that agent execution loops have become nearly real-time.

On the other hand, the Model Context Protocol (MCP) pushed by Anthropic takes a completely different direction. MCP aims for an open, client-centric protocol, focusing on flexible interoperability with local file systems and databases. While there is network latency due to frequent message passing between the client and agent, it boasts unrivaled versatility in terms of local resource compatibility and design freedom.

Ultimately, agent developers are standing at a crossroads between centralized high-speed execution for performance maximization (PTC) and open standards for flexible integration (MCP). PTC will be a powerful alternative in environments where latency is extremely critical, such as web scraping or simple API execution, but MCP still seems like an attractive option for environments that need to integrate numerous enterprise internal systems and data. haha