Augment Dual-Model Prompt

Overview (confidence: high)

Augment Code is unique in serving different prompts to different underlying models: Sonnet 4 and GPT-5. It uses a client-server SDK architecture where the remote API composes the prompt dynamically based on the model and task. Instructions are JSON-embedded with code-like function signatures for tools.

Dual-Model Architecture (confidence: high)

Augment maintains separate prompts for each underlying model:

• Sonnet 4 prompt: Optimized for Claude Sonnet 4’s capabilities and behavior patterns
• GPT-5 prompt: Optimized for GPT-5’s different architecture and strengths

This enables model-specific optimization — each prompt tailors instructions, tool descriptions, and workflow patterns to the specific model’s strengths and weaknesses. The trade-off is coordination complexity: both prompts must maintain behavioral consistency for the user.

Envelope Format (confidence: high)

JSON-embedded instructions with code-like function signatures for tools (e.g., search(query: string, max_results: int) -> Result[]). This is one of only two JSON-embedded formats in the corpus (alongside VSCode Agent/Copilot).

Architecture Pattern (confidence: high)

Client-server SDK model — the prompt lives on a remote API server. The client sends user messages and tool definitions, and the server composes the full prompt dynamically. This enables centralized prompt management: updates apply to all clients immediately.

See multi-file-agent-architecture for comparison with other architectural patterns, and monolithic-prompt-limits for the scaling challenges that motivate distributed approaches.

Implications (confidence: high)

Augment’s dual-model approach is unique in serving different prompts to different underlying models. This enables model-specific optimization but adds coordination complexity. The SDK architecture centralizes prompt management while allowing per-model specialization.

Sources

• [raw/prompts/articles/taxonomy-synthesis-2026-07-16.md]