LLM‑based agents can tackle complex multi‑step tasks, yet sequential execution incurs substantial latency. In theory, distributing work across multiple agents should yield near‑linear speedups, but existing parallel multi‑agent systems often run slower than a single‑agent baseline. We attribute this gap to two hidden costs inherent to parallel execution. The first is re‑exploration cost: parallel workers redundantly reconstruct context already held by the orchestrator (e.g., prior decisions) that would be implicitly inherited in serial execution. The second is alignment cost: extra overhead required to reconcile inconsistencies among independently generated outputs. From this we derive a principled decision criterion: a layer should be parallelized only when its critical‑path cost plus re‑exploration and alignment overheads is lower than the corresponding serial cost. This criterion is naturally expressed in wall‑clock time, but experiments show LLMs are poorly calibrated for estimating task duration. Therefore we switch to measuring cost in predicted output tokens, which LLMs estimate far more reliably. Building on this token‑based criterion, we propose SquidAgent. It estimates all token budgets in a single planning step, forks each worker directly from the orchestrator’s session to eliminate re‑exploration cost, and replaces post‑hoc reconciliation with a pre‑generated shared convention block that converts alignment into a bounded upfront cost. A deterministic scheduler then applies the criterion layer by layer. Empirically, SquidAgent achieves a $2.2\times$ mean throughput improvement and a $2.6\times$ mean wall‑time speedup over Claude Code, and a $2.0\times$ throughput gain over the strongest multi‑agent baseline.
Review