ROADMAP-23 — P3 — Parallel tool execution with bounded concurrency #115

Series: ROADMAP — filed from the 2026-10-27 whole-project assessment · Backlog task: M17-T2 · Priority: P3 · Effort: medium

Problem

Multiple tool calls in one turn are executed strictly sequentially (agentkit/core/agent.py:330–392), even though the Gemini adapter explicitly repairs/serializes multi-call histories for its provider (agentkit/llm/gemini.py:109–112) — the loop re-serializes what the model parallelized.

Acceptance criteria

• Independent tool calls in one turn dispatched via bounded asyncio.gather (new max_parallel_tools config, default sensible + documented)
• Traces remain deterministic per call id (order by call id, not completion)
• Per-tool timeout semantics unchanged; duplicate-guard semantics explicitly defined for parallel batches
• Integration test with FakeLLMClient emitting a two-call turn

Suggested branch / commit

feat/parallel-tools · feat(core): bounded parallel tool execution