Three Evolution Paths for Deterministic Software in the AI Agent Era

The author analyzes how traditional deterministic software will evolve to accommodate AI agents, identifying three primary transformation directions.

1. AI-Friendly Interaction Transformation

This involves redesigning interfaces for AI consumption rather than human users. Two aspects are highlighted:

AI-friendly UI : Interfaces previously designed for humans will be adapted for AI agents to perceive and operate.

AI-friendly API : Traditional software APIs, once called by other deterministic software, will be exposed for agent invocation. The author references emerging standards such as MCP (Model Context Protocol) and A2A (Agent-to-Agent) specifications. New agent software layers will handle orchestration and interaction, calling deterministic software components underneath.

2. AI-Friendly Runtime Transformation

Optimizations at the system level to better support inference engines and agent frameworks. Examples include memory management improvements. This falls under the System for AI category — modifying the underlying runtime to serve AI workloads more efficiently.

3. Adding Adaptiveness to Deterministic Software

Deterministic software gains adaptive capabilities by invoking AI software internally for automated decision-making and code generation. This requires a high-security, high-reliability sandbox environment to prevent safety and reliability risks. The author cites specific sandbox technologies: eBPF sandbox Wasm (WebAssembly) sandbox

Process-level security sandboxes (e.g., browser sandboxes)

This direction belongs to the AI for System category — using AI to enhance traditional software behavior.

Architectural Outlook

The author argues that the core architecture of deterministic software will not undergo massive restructuring in the near term. Instead, changes will center on AI-friendly adaptation and improved self-evolution capabilities. The primary mechanism will be protocol-based integration, likely producing adaptation or wrapper layers rather than fundamental architectural decomposition.

Application Scenario Perspective

Agent software will broaden the scope of software applicability. The article quotes an unnamed expert:

In the future, all positions' work will be software work. Large amounts of 'one-time software' (JIT software) and 'script software' are created by secretarial positions. Future agent office software will have the same status as today's Office suite; people who cannot use it will not find jobs.

The author finds this viewpoint plausible, suggesting agent-driven office tools will become as ubiquitous and essential as current productivity suites.