Exploratory HTTP Framework Benchmark

Date: 2026-07-23

Scope

This benchmark compares Agentix, Express, and NestJS-on-Express only at an identical HTTP boundary. It does not measure the preregistered coding-agent maintenance hypothesis and does not replace the Agentix-versus-plain corpus.

Each target starts a real loopback server and exposes POST /echo:

• { "value": 7 } returns HTTP 200 and the same success envelope;
• malformed JSON or any other shape returns HTTP 400 and the same validation envelope.

The runner starts all three targets outside hot-request timers, assigns one keep-alive client to each target, consumes each complete response, and rotates order using seeded blocks that contain all three stacks. Cold-ready and memory observations run in fresh child processes that dynamically import only the selected target.

Frozen exploratory run

seed: agentix-http-frameworks-exploratory-v1-2026-07-23
warmups per HTTP metric and stack: 10
measured observations per HTTP metric and stack: 100
fresh child-process observations per stack: 5

Metrics:

• valid loopback request latency;
• invalid loopback request latency;
• fresh-child dynamic-import-to-listening readiness;
• ready-process RSS;
• ready-process maximum RSS.

Raw measurements, failures, environment identity, dependency versions, Git state, lockfile hash, and source hash are retained. Summaries use R-7 quartiles, median, IQR, p95, and range.

Interpretation limits

This measures configured stacks, not abstract framework cores. NestJS uses Express underneath. Network loopback and JSON costs can dominate small differences, and a one-route microbenchmark cannot predict a non-trivial application’s performance or an AI agent’s maintenance efficiency. All results are exploratory and confirmatory-ineligible.

Results

The frozen exploratory run completed with 705 raw samples and zero unavailable measurements on Node 24.16.0, macOS arm64, an Apple M4 Max (16 logical CPUs), and 48 GiB of memory.

Express had the lowest median hot-request latency. Agentix’s median valid and invalid latency was respectively about 56% and 51% higher than Express, while NestJS-on-Express was between them. Agentix had the lowest cold-ready time and ready memory: Express used about 17% more ready RSS and NestJS about 49% more; NestJS cold readiness was about 3.5 times Agentix’s.

These are observations for this echo configuration, not a general framework ranking. In particular, the Agentix target performs its declared command dispatch and schema validation, while the Express and Nest targets use the minimal finite-number check required by the endpoint contract. The host was not dedicated or process-isolated, so unrelated background activity may have added noise despite the balanced request order.

Raw evidence:

• http-frameworks-exploratory-v1-2026-07-23.json
• raw file SHA-256: d6ea63a0a57100a62ce810ed4c828d2c19e0e4f546e2a7e32e6f541d26d56ca4
• measurement-plan SHA-256: c90319d39da34a8872e986a66c632718a52278da883e7acfb22abecaf94c1664
• comparison-source SHA-256: efcb6db8dfb4221f071488fc17adb80e3ba5bc594a79d3367f5b3bcf00a09d08
• package-lock SHA-256: 855bc97725ce43b3dbbfae4b7fa63f4ebd12c8365783a8654e86742769f46085

The report records Git commit 9ab2948002193426d9f0b0eb6d4d0ba8b87ed60a with a dirty worktree because the new comparison source had not been committed. The comparison-source and lock hashes bind the exact measured implementation and dependencies, but a clean committed rerun would be stronger evidence.

Methodology v2 run (2026-07-26)

This section records the v2 rerun after the framework’s v2 redesign. It supersedes the v1 observations above for current-code questions; the v1 record stays frozen and unedited. Classification remains exploratory; nothing here is confirmatory evidence for the agent-maintenance hypothesis.

Method changes from v1, all recorded in each report’s methodology block:

• The Agentix target uses the v2 API served by the raw serveNode host, and the v1 target’s redundant response re-validation was removed.
• A second measured condition validated gives the Express and NestJS targets zod input+output validation, matching the work Agentix always performs. The default condition keeps the v1 minimal-check targets.
• A new --isolated mode runs every target in its own child process with the client in the parent over real loopback sockets. The v1 in-process method (all three servers plus the client on one event loop) inflated cross-stack gaps roughly tenfold; the in-process mode is retained only for continuity.
• Workloads: echo POST valid/invalid (as v1), plus param-get (GET /items/:id) and echo-batch (8 in-flight keep-alive requests).
• New seed agentix-http-frameworks-exploratory-v2-2026-07-26; report schemaVersion 2 with per-sample workload/condition labels.

The current records (-r2) were produced from clean committed tree 7a36e80a4 (dirty: false) after the final-review fixes (which restored production-mode effect-output re-parsing), Node v24.16.0, Apple M4 Max, express 5.2.1, @nestjs/core 11.1.28, zod 4.4.3; 10 warmups, 300 measured iterations per workload. The initial v2 records from pre-fix tree 65ac537 are retained below as superseded linked records.

Isolated-mode medians, r2 (microseconds; agentix / express / nestjs):

Process metrics (isolated mode, default condition): cold-ready median 37.9 ms / 74.8 ms / 150.7 ms; ready RSS median 61.6 MiB / 79.7 MiB / 99.5 MiB.

In-process-mode medians follow the same ordering in every cell (for example echo-valid default 80.6 / 85.3 / 94.5). Under equal validation work (validated), Agentix’s echo-valid median is 16.2% below Express and 21.8% below NestJS in isolated mode.

Observations, subject to the same interpretation limits as v1: Agentix’s median latency was the lowest of the three stacks in every workload, in both conditions and both modes, with the lowest cold-ready time and ready RSS. These remain microbenchmarks of small handlers on one machine, not a general framework ranking.

Raw evidence (append-only; the v1 files above are unchanged):

• CURRENT: http-frameworks-exploratory-v2-isolated-2026-07-26-r2.json — SHA-256 1f609dd4bd0a39ecaef6996fcb9d7cbac87103b73b27c97611289c3ea945513e
• CURRENT: http-frameworks-exploratory-v2-inprocess-2026-07-26-r2.json — SHA-256 d4f382d812d0e1e9df352ae3c85d1fe62dcf0b0ae9a58d10f05c4e7c3844147b (fresh-process probes skipped via --no-process, recorded as unavailable)
• superseded by r2 (pre-fix tree 65ac537): http-frameworks-exploratory-v2-isolated-2026-07-26.json — SHA-256 aa3fabd63b9c886cf5486c0901baaf750905505a3205e0a66283c79acfdef30b; http-frameworks-exploratory-v2-inprocess-2026-07-26.json — SHA-256 26b8e9a0d1bb3905299d4e377d5c2f6c950dbd9c2ad4fbda09c5ac5c721ed9a5
• measurement-plan SHA-256 9d293e9015be4258984fe27e8e03ed19b7e004a9ce41672846d0e84fdbf45e71
• comparison-source SHA-256 03d2a073c02cf342d261fb63913fe336af4add47d728b9312935292269df5332
• package-lock SHA-256 896eccaf998cc6efc6bede7177c142ea4287d3cc1cca451e603804050bd963eb