ADR 0006: Model-Agnostic Provider Abstraction

Date: 2026-06-22 Status: Accepted

Context

Agent Arena needs to fairly compare different multi-agent architectures. If architectures are bound to a specific LLM provider, performance differences might stem from the model's idiosyncratic quirks rather than the architecture itself.

Decision

We abstract the LLM integration behind a ModelProvider interface that exposes a single send_message(messages, tools, system_prompt) method returning a NormalizedResponse. - Tools expose a standard JSON Schema instead of provider-specific dicts. - AnthropicProvider and GeminiProvider translate our normalized conversation history into their native formats.

Parallel Tool Call Resolution Note: During parallel tool calls, Anthropic safely matches results to calls using a unique tool_use_id for each block, meaning order is loosely required. Gemini, however, matches FunctionResponse to FunctionCall strictly by function name without unique instance IDs. If the same function is called twice in parallel, Gemini relies on the sequential order of the responses matching the original calls. Thus, the provider abstraction guarantees that tool results are appended to the normalized history in the exact order the calls were received so GeminiProvider safely generates the payload.

Consequences

• Positive: The abstraction supports multi-provider evaluation. The Phase 4 sweep ran exclusively on gemini-2.5-flash (pinned 2026-06-27) to isolate architectural variance from model variance — a deliberate design choice, not a limitation of the abstraction itself.
• Future work: Cross-model replication (e.g. Claude Sonnet) is noted as future work in ROADMAP.md and results/sweep_20260627/report.md. The abstraction is already in place to support this; no architecture changes would be required.
• Negative: Increased complexity in maintaining a two-way translation layer for tool schemas and conversation histories.