The Hidden Infrastructure Tax in Multi-Agent Coding: An Estimated Replay of 39,526 Calls and an RTX 5090 Lab

Detailed Analysis & Field Observations

When multi-agent engineering workflows feel sluggish—when a 20-turn feature session across Claude Code, Codex, and Gemini/Antigravity drags on for minutes between edits—the default assumption is almost universal: the frontier LLM is slow. Engineers respond by shopping for faster cloud endpoints or trimming system prompts. To test that assumption scientifically, we instrumented our entire production multi-LLM engineering stack, audited 39,526 raw telemetry records (121.82 hours of cumulative summed call duration across concurrent agent lanes and background workers over 14 days, including 29.92 summed hours on October 12 alone), and executed a statistically powered, 6-arm controlled laboratory study (WF-GPU-LAB-001 v3_optimized) on a local NVIDIA GeForce RTX 5090 (32 GB VRAM, Blackwell compute capability sm_120) + AMD Ryzen 7 9800X3D workstation.

What we found flips the conventional wisdom on its head. The frontier reasoning model was not the primary bottleneck. Instead, interactive velocity was being strangled by three systemic taxes: (1) Death by a Thousand Cuts from synchronous per-turn and per-tool hook spawns (~1,696 prompts and ~15,000 tool calls/week in our 7-day Oct 8-15 window) and serial test/stop gates (-o addopts=, -p no:xdist, cold interpreter imports) stealing 11 to 24+ hours/week; (2) the Background Contention Fallacy, where hidden CREATE_NO_WINDOW cloud-LLM prefetch and subagent transcript ingestion jobs ran a background storm (48.8% of all 39,526 logged calls in the 14-day Oct 2-15 ledger hit the >=29.5s timeout cap); and (3) Cold Subprocess Tax on every iterative pytest run (664.07 ms p50 just to import Python, FastAPI, and Pydantic).

Crucially, separating software fixes from hardware acceleration shows why systems engineering must precede GPU offloading. Of the 121.82 hours of 14-day cumulative logged call time, roughly 99% of the reduction comes from software and algorithmic fixes: filtering subagent transcripts out of background ingest (3,072 selfmem-supersede rows at 35.86 hrs and 4,107 selfmem-extract rows at 23.64 hrs), switching prefetch to ask2 --pointers (3,450 ledger rows at 15.45 hrs; 6,795 hook triggers in the 7-day window), reordering leftoff ahead of temporal_truth (3,467 rows at 16.23 hrs), deterministic regex/YAML rules, and the CPU-resident 4-worker WarmPytestDaemon on the Ryzen 7 9800X3D. Roughly 1 hour of logged time (5,960 embedding rows at 0.98 hrs, Stage-2 PHI escalations across 3,467 phi_screen rows, and freetext fallback across 102 codekit-checklist rows)—plus interactive foreground tail latency—is directly accelerated by the RTX 5090 GPU.

Across our five laboratory hypotheses, pairing deterministic code with resident RTX 5090 inference and CPU worker pools produced decisive gains. In H1 (N = 1,350 concurrent requests across C in {1, 2, 4, 8}), disabling qwen3 thinking preambles (think=false) raised multi-terminal canary routing accuracy from 99.63% to 1,350/1,350 (100.0%, Wilson 95% CI [99.72, 100.0]%) with zero cross-talk. In H2 (N = 30 paired trials over 5,600 768-dim vectors), warm RTX 5090 batch embedding plus float32 BLAS matrix multiplication achieved an 8.42x end-to-end speedup (77.42 ms vs 656.82 ms, Wilcoxon p = 1.86e-09).

In H3A (HIPAA PHI pre-screening across N = 100 stratified synthetic prompts), our local hardened regex (Arm 0+, 100.0% in-sample accuracy at 0.03 ms p50) and Two-Stage Cascade (Arm 4: Recall Regex -> Warm RTX 5090 qwen3:4b, 100.0% in-sample accuracy with 0.04 ms negative p50 and 88% Stage-1 fast exit, exact McNemar p = 0.015625 vs baseline Arm 0 at 93.0%) both matched 100/100, with the regex rules tuned in-sample on that same synthetic set. The CEO approved the local path on 2026-10-15: the two-stage local cascade (regex, then warm qwen3:4b) is now the default PHI admission check, and Claude Haiku (claude -p --model haiku, covered under our BAA) is the fallback when the local model is unavailable or its answer does not parse. On the live code path the local classifier answered 96 of the same 100 synthetic cases with 0 false negatives and 0 false positives at a 393 ms median; an out-of-sample check is still planned. In H3B (Brief-to-Checklist extraction across N = 50 real briefs with 370 oracle items), standalone LLMs hallucinated unverified symbols (37.07%-55.52% Micro F1), and baseline regex scored 59.98% F1, whereas Arm 0+ (enhanced AST + repo-existence filter) reached 95.35% Micro F1 in 0.50 ms and Arm 4 Hybrid (Arm 0+ unioned with repo-filtered qwen3:14b) reached 99.46% in-sample Micro F1 (367/370 items, 99.73% precision, McNemar p = 2.65e-22 vs baseline). Because both Arm 0+ rule sets were refined after error analysis on these corpora, prospective holdout validation is required to confirm out-of-sample generalization.

In H4 (N = 90 paired trials across 1, 3, and 5 real unit test files), upgrading the CPU-resident WarmPytestDaemon on the Ryzen 7 9800X3D from single-worker serial dispatch (v2, 5.00x speedup) to a 4-worker parallel file-sharded pool with mtime_ns-aware module caching (v3) delivered a 10.36x pooled speedup over cold pytest (63.15 ms vs 664.07 ms p50, Wilcoxon p = 1.74e-16), running 5-file / 74-test suites in 80.80 ms (9.43x faster than cold and 2.85x faster than v2 serial warm) with 0/150 cross-talk. Replaying all 39,526 historical telemetry rows from October 2-15, 2026—modeling both eliminated/bypassed calls and accelerated local calls—reduced cumulative 14-day subsystem duration from 121.82 hours down to an estimated 36.59 minutes in Arm 0+ Fast Mode or 55.36 minutes in Arm 4 Hybrid 14B Mode (ESTIMATED TIME). To operationalize this on a shared workstation without permanently occupying VRAM, we wired a Work-Mode ON / Gaming-Mode OFF lifecycle into services/companion/launch_portal.ps1, added scripts/warm_pytest_daemon.py, and consolidated all three PostToolUse repo hooks into .claude/hooks/post_tool_bundle.py.